xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_uvd.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: amdgpu_uvd.c,v 1.7 2020/02/14 14:34:58 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 1.7 2020/02/14 14:34:58 maya Exp $");
35 
36 #include <linux/firmware.h>
37 #include <linux/module.h>
38 #include <drm/drmP.h>
39 #include <drm/drm.h>
40 
41 #include "amdgpu.h"
42 #include "amdgpu_pm.h"
43 #include "amdgpu_uvd.h"
44 #include "cikd.h"
45 #include "uvd/uvd_4_2_d.h"
46 
47 #include <linux/nbsd-namespace.h>
48 
49 /* 1 second timeout */
50 #define UVD_IDLE_TIMEOUT_MS	1000
51 
52 /* Firmware Names */
53 #ifdef CONFIG_DRM_AMDGPU_CIK
54 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
55 #define FIRMWARE_KABINI 	"radeon/kabini_uvd.bin"
56 #define FIRMWARE_KAVERI 	"radeon/kaveri_uvd.bin"
57 #define FIRMWARE_HAWAII 	"radeon/hawaii_uvd.bin"
58 #define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
59 #endif
60 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
61 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
62 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
63 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
64 
65 /**
66  * amdgpu_uvd_cs_ctx - Command submission parser context
67  *
68  * Used for emulating virtual memory support on UVD 4.2.
69  */
70 struct amdgpu_uvd_cs_ctx {
71 	struct amdgpu_cs_parser *parser;
72 	unsigned reg, count;
73 	unsigned data0, data1;
74 	unsigned idx;
75 	unsigned ib_idx;
76 
77 	/* does the IB has a msg command */
78 	bool has_msg_cmd;
79 
80 	/* minimum buffer sizes */
81 	unsigned *buf_sizes;
82 };
83 
84 #ifdef CONFIG_DRM_AMDGPU_CIK
85 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
86 MODULE_FIRMWARE(FIRMWARE_KABINI);
87 MODULE_FIRMWARE(FIRMWARE_KAVERI);
88 MODULE_FIRMWARE(FIRMWARE_HAWAII);
89 MODULE_FIRMWARE(FIRMWARE_MULLINS);
90 #endif
91 MODULE_FIRMWARE(FIRMWARE_TONGA);
92 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
93 MODULE_FIRMWARE(FIRMWARE_FIJI);
94 MODULE_FIRMWARE(FIRMWARE_STONEY);
95 
96 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
97 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
98 
99 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
100 {
101 	unsigned long bo_size;
102 	const char *fw_name;
103 	const struct common_firmware_header *hdr;
104 	unsigned version_major, version_minor, family_id;
105 	int i, r;
106 
107 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
108 
109 	switch (adev->asic_type) {
110 #ifdef CONFIG_DRM_AMDGPU_CIK
111 	case CHIP_BONAIRE:
112 		fw_name = FIRMWARE_BONAIRE;
113 		break;
114 	case CHIP_KABINI:
115 		fw_name = FIRMWARE_KABINI;
116 		break;
117 	case CHIP_KAVERI:
118 		fw_name = FIRMWARE_KAVERI;
119 		break;
120 	case CHIP_HAWAII:
121 		fw_name = FIRMWARE_HAWAII;
122 		break;
123 	case CHIP_MULLINS:
124 		fw_name = FIRMWARE_MULLINS;
125 		break;
126 #endif
127 	case CHIP_TONGA:
128 		fw_name = FIRMWARE_TONGA;
129 		break;
130 	case CHIP_FIJI:
131 		fw_name = FIRMWARE_FIJI;
132 		break;
133 	case CHIP_CARRIZO:
134 		fw_name = FIRMWARE_CARRIZO;
135 		break;
136 	case CHIP_STONEY:
137 		fw_name = FIRMWARE_STONEY;
138 		break;
139 	default:
140 		return -EINVAL;
141 	}
142 
143 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
144 	if (r) {
145 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
146 			fw_name);
147 		return r;
148 	}
149 
150 	r = amdgpu_ucode_validate(adev->uvd.fw);
151 	if (r) {
152 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
153 			fw_name);
154 		release_firmware(adev->uvd.fw);
155 		adev->uvd.fw = NULL;
156 		return r;
157 	}
158 
159 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
160 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
161 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
162 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
163 	DRM_INFO("Found UVD firmware Version: %x.%x Family ID: %x\n",
164 		version_major, version_minor, family_id);
165 
166 	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
167 				(family_id << 8));
168 
169 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
170 		 +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
171 	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
172 			     AMDGPU_GEM_DOMAIN_VRAM,
173 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
174 			     NULL, NULL, &adev->uvd.vcpu_bo);
175 	if (r) {
176 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
177 		return r;
178 	}
179 
180 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
181 	if (r) {
182 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
183 		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
184 		return r;
185 	}
186 
187 	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
188 			  &adev->uvd.gpu_addr);
189 	if (r) {
190 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
191 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
192 		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
193 		return r;
194 	}
195 
196 	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
197 	if (r) {
198 		dev_err(adev->dev, "(%d) UVD map failed\n", r);
199 		return r;
200 	}
201 
202 	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
203 
204 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
205 		atomic_set(&adev->uvd.handles[i], 0);
206 		adev->uvd.filp[i] = NULL;
207 	}
208 
209 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
210 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
211 		adev->uvd.address_64_bit = true;
212 
213 	return 0;
214 }
215 
216 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
217 {
218 	int r;
219 
220 	if (adev->uvd.vcpu_bo == NULL)
221 		return 0;
222 
223 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
224 	if (!r) {
225 		amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
226 		amdgpu_bo_unpin(adev->uvd.vcpu_bo);
227 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
228 	}
229 
230 	amdgpu_bo_unref(&adev->uvd.vcpu_bo);
231 
232 	amdgpu_ring_fini(&adev->uvd.ring);
233 
234 	release_firmware(adev->uvd.fw);
235 
236 	return 0;
237 }
238 
239 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
240 {
241 	struct amdgpu_ring *ring = &adev->uvd.ring;
242 	int i, r;
243 
244 	if (adev->uvd.vcpu_bo == NULL)
245 		return 0;
246 
247 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
248 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
249 		if (handle != 0) {
250 			struct fence *fence;
251 
252 			amdgpu_uvd_note_usage(adev);
253 
254 			r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
255 			if (r) {
256 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
257 				continue;
258 			}
259 
260 			fence_wait(fence, false);
261 			fence_put(fence);
262 
263 			adev->uvd.filp[i] = NULL;
264 			atomic_set(&adev->uvd.handles[i], 0);
265 		}
266 	}
267 
268 	return 0;
269 }
270 
271 int amdgpu_uvd_resume(struct amdgpu_device *adev)
272 {
273 	unsigned size;
274 	void *ptr;
275 	const struct common_firmware_header *hdr;
276 	unsigned offset;
277 
278 	if (adev->uvd.vcpu_bo == NULL)
279 		return -EINVAL;
280 
281 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
282 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
283 	memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
284 		(adev->uvd.fw->size) - offset);
285 
286 	cancel_delayed_work_sync(&adev->uvd.idle_work);
287 
288 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
289 	size -= le32_to_cpu(hdr->ucode_size_bytes);
290 	ptr = adev->uvd.cpu_addr;
291 	ptr += le32_to_cpu(hdr->ucode_size_bytes);
292 
293 	memset(ptr, 0, size);
294 
295 	return 0;
296 }
297 
298 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
299 {
300 	struct amdgpu_ring *ring = &adev->uvd.ring;
301 	int i, r;
302 
303 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
304 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
305 		if (handle != 0 && adev->uvd.filp[i] == filp) {
306 			struct fence *fence;
307 
308 			amdgpu_uvd_note_usage(adev);
309 
310 			r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
311 			if (r) {
312 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
313 				continue;
314 			}
315 
316 			fence_wait(fence, false);
317 			fence_put(fence);
318 
319 			adev->uvd.filp[i] = NULL;
320 			atomic_set(&adev->uvd.handles[i], 0);
321 		}
322 	}
323 }
324 
325 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
326 {
327 	int i;
328 	for (i = 0; i < rbo->placement.num_placement; ++i) {
329 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
330 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
331 	}
332 }
333 
334 /**
335  * amdgpu_uvd_cs_pass1 - first parsing round
336  *
337  * @ctx: UVD parser context
338  *
339  * Make sure UVD message and feedback buffers are in VRAM and
340  * nobody is violating an 256MB boundary.
341  */
342 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
343 {
344 	struct amdgpu_bo_va_mapping *mapping;
345 	struct amdgpu_bo *bo;
346 	uint32_t cmd, lo, hi;
347 	uint64_t addr;
348 	int r = 0;
349 
350 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
351 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
352 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
353 
354 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
355 	if (mapping == NULL) {
356 		DRM_ERROR("Can't find BO for addr 0x%08"PRIx64"\n", addr);
357 		return -EINVAL;
358 	}
359 
360 	if (!ctx->parser->adev->uvd.address_64_bit) {
361 		/* check if it's a message or feedback command */
362 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
363 		if (cmd == 0x0 || cmd == 0x3) {
364 			/* yes, force it into VRAM */
365 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
366 			amdgpu_ttm_placement_from_domain(bo, domain);
367 		}
368 		amdgpu_uvd_force_into_uvd_segment(bo);
369 
370 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
371 	}
372 
373 	return r;
374 }
375 
376 /**
377  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
378  *
379  * @msg: pointer to message structure
380  * @buf_sizes: returned buffer sizes
381  *
382  * Peek into the decode message and calculate the necessary buffer sizes.
383  */
384 static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
385 {
386 	unsigned stream_type = msg[4];
387 	unsigned width = msg[6];
388 	unsigned height = msg[7];
389 	unsigned dpb_size = msg[9];
390 	unsigned pitch = msg[28];
391 	unsigned level = msg[57];
392 
393 	unsigned width_in_mb = width / 16;
394 	unsigned height_in_mb = ALIGN(height / 16, 2);
395 	unsigned fs_in_mb = width_in_mb * height_in_mb;
396 
397 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
398 	unsigned min_ctx_size = 0;
399 
400 	image_size = width * height;
401 	image_size += image_size / 2;
402 	image_size = ALIGN(image_size, 1024);
403 
404 	switch (stream_type) {
405 	case 0: /* H264 */
406 	case 7: /* H264 Perf */
407 		switch(level) {
408 		case 30:
409 			num_dpb_buffer = 8100 / fs_in_mb;
410 			break;
411 		case 31:
412 			num_dpb_buffer = 18000 / fs_in_mb;
413 			break;
414 		case 32:
415 			num_dpb_buffer = 20480 / fs_in_mb;
416 			break;
417 		case 41:
418 			num_dpb_buffer = 32768 / fs_in_mb;
419 			break;
420 		case 42:
421 			num_dpb_buffer = 34816 / fs_in_mb;
422 			break;
423 		case 50:
424 			num_dpb_buffer = 110400 / fs_in_mb;
425 			break;
426 		case 51:
427 			num_dpb_buffer = 184320 / fs_in_mb;
428 			break;
429 		default:
430 			num_dpb_buffer = 184320 / fs_in_mb;
431 			break;
432 		}
433 		num_dpb_buffer++;
434 		if (num_dpb_buffer > 17)
435 			num_dpb_buffer = 17;
436 
437 		/* reference picture buffer */
438 		min_dpb_size = image_size * num_dpb_buffer;
439 
440 		/* macroblock context buffer */
441 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
442 
443 		/* IT surface buffer */
444 		min_dpb_size += width_in_mb * height_in_mb * 32;
445 		break;
446 
447 	case 1: /* VC1 */
448 
449 		/* reference picture buffer */
450 		min_dpb_size = image_size * 3;
451 
452 		/* CONTEXT_BUFFER */
453 		min_dpb_size += width_in_mb * height_in_mb * 128;
454 
455 		/* IT surface buffer */
456 		min_dpb_size += width_in_mb * 64;
457 
458 		/* DB surface buffer */
459 		min_dpb_size += width_in_mb * 128;
460 
461 		/* BP */
462 		tmp = max(width_in_mb, height_in_mb);
463 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
464 		break;
465 
466 	case 3: /* MPEG2 */
467 
468 		/* reference picture buffer */
469 		min_dpb_size = image_size * 3;
470 		break;
471 
472 	case 4: /* MPEG4 */
473 
474 		/* reference picture buffer */
475 		min_dpb_size = image_size * 3;
476 
477 		/* CM */
478 		min_dpb_size += width_in_mb * height_in_mb * 64;
479 
480 		/* IT surface buffer */
481 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
482 		break;
483 
484 	case 16: /* H265 */
485 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
486 		image_size = ALIGN(image_size, 256);
487 
488 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
489 		min_dpb_size = image_size * num_dpb_buffer;
490 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
491 					   * 16 * num_dpb_buffer + 52 * 1024;
492 		break;
493 
494 	default:
495 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
496 		return -EINVAL;
497 	}
498 
499 	if (width > pitch) {
500 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
501 		return -EINVAL;
502 	}
503 
504 	if (dpb_size < min_dpb_size) {
505 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
506 			  dpb_size, min_dpb_size);
507 		return -EINVAL;
508 	}
509 
510 	buf_sizes[0x1] = dpb_size;
511 	buf_sizes[0x2] = image_size;
512 	buf_sizes[0x4] = min_ctx_size;
513 	return 0;
514 }
515 
516 /**
517  * amdgpu_uvd_cs_msg - handle UVD message
518  *
519  * @ctx: UVD parser context
520  * @bo: buffer object containing the message
521  * @offset: offset into the buffer object
522  *
523  * Peek into the UVD message and extract the session id.
524  * Make sure that we don't open up to many sessions.
525  */
526 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
527 			     struct amdgpu_bo *bo, unsigned offset)
528 {
529 	struct amdgpu_device *adev = ctx->parser->adev;
530 	int32_t *msg, msg_type, handle;
531 	void *ptr;
532 	long r;
533 	int i;
534 
535 	if (offset & 0x3F) {
536 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
537 		return -EINVAL;
538 	}
539 
540 	r = reservation_object_wait_timeout_rcu(bo->tbo.resv, true, false,
541 						MAX_SCHEDULE_TIMEOUT);
542 	if (r < 0) {
543 		DRM_ERROR("Failed waiting for UVD message (%ld)!\n", r);
544 		return r;
545 	}
546 
547 	r = amdgpu_bo_kmap(bo, &ptr);
548 	if (r) {
549 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
550 		return r;
551 	}
552 
553 	msg = ptr + offset;
554 
555 	msg_type = msg[1];
556 	handle = msg[2];
557 
558 	if (handle == 0) {
559 		DRM_ERROR("Invalid UVD handle!\n");
560 		return -EINVAL;
561 	}
562 
563 	switch (msg_type) {
564 	case 0:
565 		/* it's a create msg, calc image size (width * height) */
566 		amdgpu_bo_kunmap(bo);
567 
568 		/* try to alloc a new handle */
569 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
570 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
571 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
572 				return -EINVAL;
573 			}
574 
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 	case 1:
585 		/* it's a decode msg, calc buffer sizes */
586 		r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
587 		amdgpu_bo_kunmap(bo);
588 		if (r)
589 			return r;
590 
591 		/* validate the handle */
592 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
593 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
594 				if (adev->uvd.filp[i] != ctx->parser->filp) {
595 					DRM_ERROR("UVD handle collision detected!\n");
596 					return -EINVAL;
597 				}
598 				return 0;
599 			}
600 		}
601 
602 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
603 		return -ENOENT;
604 
605 	case 2:
606 		/* it's a destroy msg, free the handle */
607 		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
608 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
609 		amdgpu_bo_kunmap(bo);
610 		return 0;
611 
612 	default:
613 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
614 		return -EINVAL;
615 	}
616 	BUG();
617 	return -EINVAL;
618 }
619 
620 /**
621  * amdgpu_uvd_cs_pass2 - second parsing round
622  *
623  * @ctx: UVD parser context
624  *
625  * Patch buffer addresses, make sure buffer sizes are correct.
626  */
627 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
628 {
629 	struct amdgpu_bo_va_mapping *mapping;
630 	struct amdgpu_bo *bo;
631 	struct amdgpu_ib *ib;
632 	uint32_t cmd, lo, hi;
633 	uint64_t start, end;
634 	uint64_t addr;
635 	int r;
636 
637 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
638 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
639 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
640 
641 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
642 	if (mapping == NULL)
643 		return -EINVAL;
644 
645 	start = amdgpu_bo_gpu_offset(bo);
646 
647 	end = (mapping->it.last + 1 - mapping->it.start);
648 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
649 
650 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
651 	start += addr;
652 
653 	ib = &ctx->parser->ibs[ctx->ib_idx];
654 	ib->ptr[ctx->data0] = start & 0xFFFFFFFF;
655 	ib->ptr[ctx->data1] = start >> 32;
656 
657 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
658 	if (cmd < 0x4) {
659 		if ((end - start) < ctx->buf_sizes[cmd]) {
660 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
661 				  (unsigned)(end - start),
662 				  ctx->buf_sizes[cmd]);
663 			return -EINVAL;
664 		}
665 
666 	} else if (cmd == 0x206) {
667 		if ((end - start) < ctx->buf_sizes[4]) {
668 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
669 					  (unsigned)(end - start),
670 					  ctx->buf_sizes[4]);
671 			return -EINVAL;
672 		}
673 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
674 		DRM_ERROR("invalid UVD command %X!\n", cmd);
675 		return -EINVAL;
676 	}
677 
678 	if (!ctx->parser->adev->uvd.address_64_bit) {
679 		if ((start >> 28) != ((end - 1) >> 28)) {
680 			DRM_ERROR("reloc %"PRIX64"-%"PRIX64" crossing 256MB boundary!\n",
681 				  start, end);
682 			return -EINVAL;
683 		}
684 
685 		if ((cmd == 0 || cmd == 0x3) &&
686 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
687 			DRM_ERROR("msg/fb buffer %"PRIX64"-%"PRIX64" out of 256MB segment!\n",
688 				  start, end);
689 			return -EINVAL;
690 		}
691 	}
692 
693 	if (cmd == 0) {
694 		ctx->has_msg_cmd = true;
695 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
696 		if (r)
697 			return r;
698 	} else if (!ctx->has_msg_cmd) {
699 		DRM_ERROR("Message needed before other commands are send!\n");
700 		return -EINVAL;
701 	}
702 
703 	return 0;
704 }
705 
706 /**
707  * amdgpu_uvd_cs_reg - parse register writes
708  *
709  * @ctx: UVD parser context
710  * @cb: callback function
711  *
712  * Parse the register writes, call cb on each complete command.
713  */
714 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
715 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
716 {
717 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
718 	int i, r;
719 
720 	ctx->idx++;
721 	for (i = 0; i <= ctx->count; ++i) {
722 		unsigned reg = ctx->reg + i;
723 
724 		if (ctx->idx >= ib->length_dw) {
725 			DRM_ERROR("Register command after end of CS!\n");
726 			return -EINVAL;
727 		}
728 
729 		switch (reg) {
730 		case mmUVD_GPCOM_VCPU_DATA0:
731 			ctx->data0 = ctx->idx;
732 			break;
733 		case mmUVD_GPCOM_VCPU_DATA1:
734 			ctx->data1 = ctx->idx;
735 			break;
736 		case mmUVD_GPCOM_VCPU_CMD:
737 			r = cb(ctx);
738 			if (r)
739 				return r;
740 			break;
741 		case mmUVD_ENGINE_CNTL:
742 			break;
743 		default:
744 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
745 			return -EINVAL;
746 		}
747 		ctx->idx++;
748 	}
749 	return 0;
750 }
751 
752 /**
753  * amdgpu_uvd_cs_packets - parse UVD packets
754  *
755  * @ctx: UVD parser context
756  * @cb: callback function
757  *
758  * Parse the command stream packets.
759  */
760 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
761 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
762 {
763 	struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
764 	int r;
765 
766 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
767 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
768 		unsigned type = CP_PACKET_GET_TYPE(cmd);
769 		switch (type) {
770 		case PACKET_TYPE0:
771 			ctx->reg = CP_PACKET0_GET_REG(cmd);
772 			ctx->count = CP_PACKET_GET_COUNT(cmd);
773 			r = amdgpu_uvd_cs_reg(ctx, cb);
774 			if (r)
775 				return r;
776 			break;
777 		case PACKET_TYPE2:
778 			++ctx->idx;
779 			break;
780 		default:
781 			DRM_ERROR("Unknown packet type %d !\n", type);
782 			return -EINVAL;
783 		}
784 	}
785 	return 0;
786 }
787 
788 /**
789  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
790  *
791  * @parser: Command submission parser context
792  *
793  * Parse the command stream, patch in addresses as necessary.
794  */
795 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
796 {
797 	struct amdgpu_uvd_cs_ctx ctx = {};
798 	unsigned buf_sizes[] = {
799 		[0x00000000]	=	2048,
800 		[0x00000001]	=	0xFFFFFFFF,
801 		[0x00000002]	=	0xFFFFFFFF,
802 		[0x00000003]	=	2048,
803 		[0x00000004]	=	0xFFFFFFFF,
804 	};
805 	struct amdgpu_ib *ib = &parser->ibs[ib_idx];
806 	int r;
807 
808 	if (ib->length_dw % 16) {
809 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
810 			  ib->length_dw);
811 		return -EINVAL;
812 	}
813 
814 	ctx.parser = parser;
815 	ctx.buf_sizes = buf_sizes;
816 	ctx.ib_idx = ib_idx;
817 
818 	/* first round, make sure the buffers are actually in the UVD segment */
819 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
820 	if (r)
821 		return r;
822 
823 	/* second round, patch buffer addresses into the command stream */
824 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
825 	if (r)
826 		return r;
827 
828 	if (!ctx.has_msg_cmd) {
829 		DRM_ERROR("UVD-IBs need a msg command!\n");
830 		return -EINVAL;
831 	}
832 
833 	amdgpu_uvd_note_usage(ctx.parser->adev);
834 
835 	return 0;
836 }
837 
838 static int amdgpu_uvd_free_job(
839 	struct amdgpu_job *job)
840 {
841 	amdgpu_ib_free(job->adev, job->ibs);
842 	kfree(job->ibs);
843 	return 0;
844 }
845 
846 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring,
847 			       struct amdgpu_bo *bo,
848 			       struct fence **fence)
849 {
850 	struct ttm_validate_buffer tv;
851 	struct ww_acquire_ctx ticket;
852 	struct list_head head;
853 	struct amdgpu_ib *ib = NULL;
854 	struct fence *f = NULL;
855 	struct amdgpu_device *adev = ring->adev;
856 	uint64_t addr;
857 	int i, r;
858 
859 	memset(&tv, 0, sizeof(tv));
860 	tv.bo = &bo->tbo;
861 
862 	INIT_LIST_HEAD(&head);
863 	list_add(&tv.head, &head);
864 
865 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
866 	if (r)
867 		return r;
868 
869 	if (!bo->adev->uvd.address_64_bit) {
870 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
871 		amdgpu_uvd_force_into_uvd_segment(bo);
872 	}
873 
874 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
875 	if (r)
876 		goto err;
877 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
878 	if (!ib) {
879 		r = -ENOMEM;
880 		goto err;
881 	}
882 	r = amdgpu_ib_get(ring, NULL, 64, ib);
883 	if (r)
884 		goto err1;
885 
886 	addr = amdgpu_bo_gpu_offset(bo);
887 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
888 	ib->ptr[1] = addr;
889 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
890 	ib->ptr[3] = addr >> 32;
891 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
892 	ib->ptr[5] = 0;
893 	for (i = 6; i < 16; ++i)
894 		ib->ptr[i] = PACKET2(0);
895 	ib->length_dw = 16;
896 
897 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
898 						 &amdgpu_uvd_free_job,
899 						 AMDGPU_FENCE_OWNER_UNDEFINED,
900 						 &f);
901 	if (r)
902 		goto err2;
903 
904 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
905 
906 	if (fence)
907 		*fence = fence_get(f);
908 	amdgpu_bo_unref(&bo);
909 	fence_put(f);
910 	if (amdgpu_enable_scheduler)
911 		return 0;
912 
913 	amdgpu_ib_free(ring->adev, ib);
914 	kfree(ib);
915 	return 0;
916 err2:
917 	amdgpu_ib_free(ring->adev, ib);
918 err1:
919 	kfree(ib);
920 err:
921 	ttm_eu_backoff_reservation(&ticket, &head);
922 	return r;
923 }
924 
925 /* multiple fence commands without any stream commands in between can
926    crash the vcpu so just try to emmit a dummy create/destroy msg to
927    avoid this */
928 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
929 			      struct fence **fence)
930 {
931 	struct amdgpu_device *adev = ring->adev;
932 	struct amdgpu_bo *bo;
933 	uint32_t *msg;
934 	int r, i;
935 
936 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
937 			     AMDGPU_GEM_DOMAIN_VRAM,
938 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
939 			     NULL, NULL, &bo);
940 	if (r)
941 		return r;
942 
943 	r = amdgpu_bo_reserve(bo, false);
944 	if (r) {
945 		amdgpu_bo_unref(&bo);
946 		return r;
947 	}
948 
949 	r = amdgpu_bo_kmap(bo, (void **)&msg);
950 	if (r) {
951 		amdgpu_bo_unreserve(bo);
952 		amdgpu_bo_unref(&bo);
953 		return r;
954 	}
955 
956 	/* stitch together an UVD create msg */
957 	msg[0] = cpu_to_le32(0x00000de4);
958 	msg[1] = cpu_to_le32(0x00000000);
959 	msg[2] = cpu_to_le32(handle);
960 	msg[3] = cpu_to_le32(0x00000000);
961 	msg[4] = cpu_to_le32(0x00000000);
962 	msg[5] = cpu_to_le32(0x00000000);
963 	msg[6] = cpu_to_le32(0x00000000);
964 	msg[7] = cpu_to_le32(0x00000780);
965 	msg[8] = cpu_to_le32(0x00000440);
966 	msg[9] = cpu_to_le32(0x00000000);
967 	msg[10] = cpu_to_le32(0x01b37000);
968 	for (i = 11; i < 1024; ++i)
969 		msg[i] = cpu_to_le32(0x0);
970 
971 	amdgpu_bo_kunmap(bo);
972 	amdgpu_bo_unreserve(bo);
973 
974 	return amdgpu_uvd_send_msg(ring, bo, fence);
975 }
976 
977 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
978 			       struct fence **fence)
979 {
980 	struct amdgpu_device *adev = ring->adev;
981 	struct amdgpu_bo *bo;
982 	uint32_t *msg;
983 	int r, i;
984 
985 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
986 			     AMDGPU_GEM_DOMAIN_VRAM,
987 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
988 			     NULL, NULL, &bo);
989 	if (r)
990 		return r;
991 
992 	r = amdgpu_bo_reserve(bo, false);
993 	if (r) {
994 		amdgpu_bo_unref(&bo);
995 		return r;
996 	}
997 
998 	r = amdgpu_bo_kmap(bo, (void **)&msg);
999 	if (r) {
1000 		amdgpu_bo_unreserve(bo);
1001 		amdgpu_bo_unref(&bo);
1002 		return r;
1003 	}
1004 
1005 	/* stitch together an UVD destroy msg */
1006 	msg[0] = cpu_to_le32(0x00000de4);
1007 	msg[1] = cpu_to_le32(0x00000002);
1008 	msg[2] = cpu_to_le32(handle);
1009 	msg[3] = cpu_to_le32(0x00000000);
1010 	for (i = 4; i < 1024; ++i)
1011 		msg[i] = cpu_to_le32(0x0);
1012 
1013 	amdgpu_bo_kunmap(bo);
1014 	amdgpu_bo_unreserve(bo);
1015 
1016 	return amdgpu_uvd_send_msg(ring, bo, fence);
1017 }
1018 
1019 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1020 {
1021 	struct amdgpu_device *adev =
1022 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1023 	unsigned i, fences, handles = 0;
1024 
1025 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1026 
1027 	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1028 		if (atomic_read(&adev->uvd.handles[i]))
1029 			++handles;
1030 
1031 	if (fences == 0 && handles == 0) {
1032 		if (adev->pm.dpm_enabled) {
1033 			amdgpu_dpm_enable_uvd(adev, false);
1034 		} else {
1035 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1036 		}
1037 	} else {
1038 		schedule_delayed_work(&adev->uvd.idle_work,
1039 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1040 	}
1041 }
1042 
1043 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1044 {
1045 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1046 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1047 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1048 
1049 	if (set_clocks) {
1050 		if (adev->pm.dpm_enabled) {
1051 			amdgpu_dpm_enable_uvd(adev, true);
1052 		} else {
1053 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1054 		}
1055 	}
1056 }
1057