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 /* Polaris10/11 firmware version */
45 #define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
46 
47 /* Firmware Names */
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
50 #define FIRMWARE_KABINI	"radeon/kabini_uvd.bin"
51 #define FIRMWARE_KAVERI	"radeon/kaveri_uvd.bin"
52 #define FIRMWARE_HAWAII	"radeon/hawaii_uvd.bin"
53 #define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
54 #endif
55 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
56 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
57 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
58 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
59 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
60 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
61 
62 /**
63  * amdgpu_uvd_cs_ctx - Command submission parser context
64  *
65  * Used for emulating virtual memory support on UVD 4.2.
66  */
67 struct amdgpu_uvd_cs_ctx {
68 	struct amdgpu_cs_parser *parser;
69 	unsigned reg, count;
70 	unsigned data0, data1;
71 	unsigned idx;
72 	unsigned ib_idx;
73 
74 	/* does the IB has a msg command */
75 	bool has_msg_cmd;
76 
77 	/* minimum buffer sizes */
78 	unsigned *buf_sizes;
79 };
80 
81 #ifdef CONFIG_DRM_AMDGPU_CIK
82 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
83 MODULE_FIRMWARE(FIRMWARE_KABINI);
84 MODULE_FIRMWARE(FIRMWARE_KAVERI);
85 MODULE_FIRMWARE(FIRMWARE_HAWAII);
86 MODULE_FIRMWARE(FIRMWARE_MULLINS);
87 #endif
88 MODULE_FIRMWARE(FIRMWARE_TONGA);
89 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
90 MODULE_FIRMWARE(FIRMWARE_FIJI);
91 MODULE_FIRMWARE(FIRMWARE_STONEY);
92 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
93 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
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 	struct amdgpu_ring *ring;
101 	struct amd_sched_rq *rq;
102 	unsigned long bo_size;
103 	const char *fw_name;
104 	const struct common_firmware_header *hdr;
105 	unsigned version_major, version_minor, family_id;
106 	int i, r;
107 
108 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
109 
110 	switch (adev->asic_type) {
111 #ifdef CONFIG_DRM_AMDGPU_CIK
112 	case CHIP_BONAIRE:
113 		fw_name = FIRMWARE_BONAIRE;
114 		break;
115 	case CHIP_KABINI:
116 		fw_name = FIRMWARE_KABINI;
117 		break;
118 	case CHIP_KAVERI:
119 		fw_name = FIRMWARE_KAVERI;
120 		break;
121 	case CHIP_HAWAII:
122 		fw_name = FIRMWARE_HAWAII;
123 		break;
124 	case CHIP_MULLINS:
125 		fw_name = FIRMWARE_MULLINS;
126 		break;
127 #endif
128 	case CHIP_TONGA:
129 		fw_name = FIRMWARE_TONGA;
130 		break;
131 	case CHIP_FIJI:
132 		fw_name = FIRMWARE_FIJI;
133 		break;
134 	case CHIP_CARRIZO:
135 		fw_name = FIRMWARE_CARRIZO;
136 		break;
137 	case CHIP_STONEY:
138 		fw_name = FIRMWARE_STONEY;
139 		break;
140 	case CHIP_POLARIS10:
141 		fw_name = FIRMWARE_POLARIS10;
142 		break;
143 	case CHIP_POLARIS11:
144 		fw_name = FIRMWARE_POLARIS11;
145 		break;
146 	default:
147 		return -EINVAL;
148 	}
149 
150 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
151 	if (r) {
152 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
153 			fw_name);
154 		return r;
155 	}
156 
157 	r = amdgpu_ucode_validate(adev->uvd.fw);
158 	if (r) {
159 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
160 			fw_name);
161 		release_firmware(adev->uvd.fw);
162 		adev->uvd.fw = NULL;
163 		return r;
164 	}
165 
166 	/* Set the default UVD handles that the firmware can handle */
167 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
168 
169 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
170 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
171 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
172 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
173 	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
174 		version_major, version_minor, family_id);
175 
176 	/*
177 	 * Limit the number of UVD handles depending on microcode major
178 	 * and minor versions. The firmware version which has 40 UVD
179 	 * instances support is 1.80. So all subsequent versions should
180 	 * also have the same support.
181 	 */
182 	if ((version_major > 0x01) ||
183 	    ((version_major == 0x01) && (version_minor >= 0x50)))
184 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
185 
186 	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
187 				(family_id << 8));
188 
189 	if ((adev->asic_type == CHIP_POLARIS10 ||
190 	     adev->asic_type == CHIP_POLARIS11) &&
191 	    (adev->uvd.fw_version < FW_1_66_16))
192 		DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
193 			  version_major, version_minor);
194 
195 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
196 		  +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
197 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
198 	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
199 			     AMDGPU_GEM_DOMAIN_VRAM,
200 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
201 			     NULL, NULL, &adev->uvd.vcpu_bo);
202 	if (r) {
203 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
204 		return r;
205 	}
206 
207 	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
208 	if (r) {
209 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
210 		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
211 		return r;
212 	}
213 
214 	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
215 			  &adev->uvd.gpu_addr);
216 	if (r) {
217 		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
218 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
219 		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
220 		return r;
221 	}
222 
223 	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
224 	if (r) {
225 		dev_err(adev->dev, "(%d) UVD map failed\n", r);
226 		return r;
227 	}
228 
229 	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
230 
231 	ring = &adev->uvd.ring;
232 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
233 	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
234 				  rq, amdgpu_sched_jobs);
235 	if (r != 0) {
236 		DRM_ERROR("Failed setting up UVD run queue.\n");
237 		return r;
238 	}
239 
240 	for (i = 0; i < adev->uvd.max_handles; ++i) {
241 		atomic_set(&adev->uvd.handles[i], 0);
242 		adev->uvd.filp[i] = NULL;
243 	}
244 
245 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
246 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
247 		adev->uvd.address_64_bit = true;
248 
249 	return 0;
250 }
251 
252 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
253 {
254 	int r;
255 
256 	kfree(adev->uvd.saved_bo);
257 
258 	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
259 
260 	if (adev->uvd.vcpu_bo) {
261 		r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
262 		if (!r) {
263 			amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
264 			amdgpu_bo_unpin(adev->uvd.vcpu_bo);
265 			amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
266 		}
267 
268 		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
269 	}
270 
271 	amdgpu_ring_fini(&adev->uvd.ring);
272 
273 	release_firmware(adev->uvd.fw);
274 
275 	return 0;
276 }
277 
278 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
279 {
280 	unsigned size;
281 	void *ptr;
282 	int i;
283 
284 	if (adev->uvd.vcpu_bo == NULL)
285 		return 0;
286 
287 	for (i = 0; i < adev->uvd.max_handles; ++i)
288 		if (atomic_read(&adev->uvd.handles[i]))
289 			break;
290 
291 	if (i == AMDGPU_MAX_UVD_HANDLES)
292 		return 0;
293 
294 	cancel_delayed_work_sync(&adev->uvd.idle_work);
295 
296 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
297 	ptr = adev->uvd.cpu_addr;
298 
299 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
300 	if (!adev->uvd.saved_bo)
301 		return -ENOMEM;
302 
303 	memcpy(adev->uvd.saved_bo, ptr, size);
304 
305 	return 0;
306 }
307 
308 int amdgpu_uvd_resume(struct amdgpu_device *adev)
309 {
310 	unsigned size;
311 	void *ptr;
312 
313 	if (adev->uvd.vcpu_bo == NULL)
314 		return -EINVAL;
315 
316 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
317 	ptr = adev->uvd.cpu_addr;
318 
319 	if (adev->uvd.saved_bo != NULL) {
320 		memcpy(ptr, adev->uvd.saved_bo, size);
321 		kfree(adev->uvd.saved_bo);
322 		adev->uvd.saved_bo = NULL;
323 	} else {
324 		const struct common_firmware_header *hdr;
325 		unsigned offset;
326 
327 		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
328 		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
329 		memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
330 			(adev->uvd.fw->size) - offset);
331 		size -= le32_to_cpu(hdr->ucode_size_bytes);
332 		ptr += le32_to_cpu(hdr->ucode_size_bytes);
333 		memset(ptr, 0, size);
334 	}
335 
336 	return 0;
337 }
338 
339 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
340 {
341 	struct amdgpu_ring *ring = &adev->uvd.ring;
342 	int i, r;
343 
344 	for (i = 0; i < adev->uvd.max_handles; ++i) {
345 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
346 		if (handle != 0 && adev->uvd.filp[i] == filp) {
347 			struct fence *fence;
348 
349 			amdgpu_uvd_note_usage(adev);
350 
351 			r = amdgpu_uvd_get_destroy_msg(ring, handle,
352 						       false, &fence);
353 			if (r) {
354 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
355 				continue;
356 			}
357 
358 			fence_wait(fence, false);
359 			fence_put(fence);
360 
361 			adev->uvd.filp[i] = NULL;
362 			atomic_set(&adev->uvd.handles[i], 0);
363 		}
364 	}
365 }
366 
367 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
368 {
369 	int i;
370 	for (i = 0; i < rbo->placement.num_placement; ++i) {
371 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
372 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
373 	}
374 }
375 
376 /**
377  * amdgpu_uvd_cs_pass1 - first parsing round
378  *
379  * @ctx: UVD parser context
380  *
381  * Make sure UVD message and feedback buffers are in VRAM and
382  * nobody is violating an 256MB boundary.
383  */
384 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
385 {
386 	struct amdgpu_bo_va_mapping *mapping;
387 	struct amdgpu_bo *bo;
388 	uint32_t cmd, lo, hi;
389 	uint64_t addr;
390 	int r = 0;
391 
392 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
393 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
394 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
395 
396 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
397 	if (mapping == NULL) {
398 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
399 		return -EINVAL;
400 	}
401 
402 	if (!ctx->parser->adev->uvd.address_64_bit) {
403 		/* check if it's a message or feedback command */
404 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
405 		if (cmd == 0x0 || cmd == 0x3) {
406 			/* yes, force it into VRAM */
407 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
408 			amdgpu_ttm_placement_from_domain(bo, domain);
409 		}
410 		amdgpu_uvd_force_into_uvd_segment(bo);
411 
412 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
413 	}
414 
415 	return r;
416 }
417 
418 /**
419  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
420  *
421  * @msg: pointer to message structure
422  * @buf_sizes: returned buffer sizes
423  *
424  * Peek into the decode message and calculate the necessary buffer sizes.
425  */
426 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
427 	unsigned buf_sizes[])
428 {
429 	unsigned stream_type = msg[4];
430 	unsigned width = msg[6];
431 	unsigned height = msg[7];
432 	unsigned dpb_size = msg[9];
433 	unsigned pitch = msg[28];
434 	unsigned level = msg[57];
435 
436 	unsigned width_in_mb = width / 16;
437 	unsigned height_in_mb = ALIGN(height / 16, 2);
438 	unsigned fs_in_mb = width_in_mb * height_in_mb;
439 
440 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
441 	unsigned min_ctx_size = 0;
442 
443 	image_size = width * height;
444 	image_size += image_size / 2;
445 	image_size = ALIGN(image_size, 1024);
446 
447 	switch (stream_type) {
448 	case 0: /* H264 */
449 		switch(level) {
450 		case 30:
451 			num_dpb_buffer = 8100 / fs_in_mb;
452 			break;
453 		case 31:
454 			num_dpb_buffer = 18000 / fs_in_mb;
455 			break;
456 		case 32:
457 			num_dpb_buffer = 20480 / fs_in_mb;
458 			break;
459 		case 41:
460 			num_dpb_buffer = 32768 / fs_in_mb;
461 			break;
462 		case 42:
463 			num_dpb_buffer = 34816 / fs_in_mb;
464 			break;
465 		case 50:
466 			num_dpb_buffer = 110400 / fs_in_mb;
467 			break;
468 		case 51:
469 			num_dpb_buffer = 184320 / fs_in_mb;
470 			break;
471 		default:
472 			num_dpb_buffer = 184320 / fs_in_mb;
473 			break;
474 		}
475 		num_dpb_buffer++;
476 		if (num_dpb_buffer > 17)
477 			num_dpb_buffer = 17;
478 
479 		/* reference picture buffer */
480 		min_dpb_size = image_size * num_dpb_buffer;
481 
482 		/* macroblock context buffer */
483 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
484 
485 		/* IT surface buffer */
486 		min_dpb_size += width_in_mb * height_in_mb * 32;
487 		break;
488 
489 	case 1: /* VC1 */
490 
491 		/* reference picture buffer */
492 		min_dpb_size = image_size * 3;
493 
494 		/* CONTEXT_BUFFER */
495 		min_dpb_size += width_in_mb * height_in_mb * 128;
496 
497 		/* IT surface buffer */
498 		min_dpb_size += width_in_mb * 64;
499 
500 		/* DB surface buffer */
501 		min_dpb_size += width_in_mb * 128;
502 
503 		/* BP */
504 		tmp = max(width_in_mb, height_in_mb);
505 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
506 		break;
507 
508 	case 3: /* MPEG2 */
509 
510 		/* reference picture buffer */
511 		min_dpb_size = image_size * 3;
512 		break;
513 
514 	case 4: /* MPEG4 */
515 
516 		/* reference picture buffer */
517 		min_dpb_size = image_size * 3;
518 
519 		/* CM */
520 		min_dpb_size += width_in_mb * height_in_mb * 64;
521 
522 		/* IT surface buffer */
523 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
524 		break;
525 
526 	case 7: /* H264 Perf */
527 		switch(level) {
528 		case 30:
529 			num_dpb_buffer = 8100 / fs_in_mb;
530 			break;
531 		case 31:
532 			num_dpb_buffer = 18000 / fs_in_mb;
533 			break;
534 		case 32:
535 			num_dpb_buffer = 20480 / fs_in_mb;
536 			break;
537 		case 41:
538 			num_dpb_buffer = 32768 / fs_in_mb;
539 			break;
540 		case 42:
541 			num_dpb_buffer = 34816 / fs_in_mb;
542 			break;
543 		case 50:
544 			num_dpb_buffer = 110400 / fs_in_mb;
545 			break;
546 		case 51:
547 			num_dpb_buffer = 184320 / fs_in_mb;
548 			break;
549 		default:
550 			num_dpb_buffer = 184320 / fs_in_mb;
551 			break;
552 		}
553 		num_dpb_buffer++;
554 		if (num_dpb_buffer > 17)
555 			num_dpb_buffer = 17;
556 
557 		/* reference picture buffer */
558 		min_dpb_size = image_size * num_dpb_buffer;
559 
560 		if (adev->asic_type < CHIP_POLARIS10){
561 			/* macroblock context buffer */
562 			min_dpb_size +=
563 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
564 
565 			/* IT surface buffer */
566 			min_dpb_size += width_in_mb * height_in_mb * 32;
567 		} else {
568 			/* macroblock context buffer */
569 			min_ctx_size =
570 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
571 		}
572 		break;
573 
574 	case 16: /* H265 */
575 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
576 		image_size = ALIGN(image_size, 256);
577 
578 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
579 		min_dpb_size = image_size * num_dpb_buffer;
580 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
581 					   * 16 * num_dpb_buffer + 52 * 1024;
582 		break;
583 
584 	default:
585 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
586 		return -EINVAL;
587 	}
588 
589 	if (width > pitch) {
590 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
591 		return -EINVAL;
592 	}
593 
594 	if (dpb_size < min_dpb_size) {
595 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
596 			  dpb_size, min_dpb_size);
597 		return -EINVAL;
598 	}
599 
600 	buf_sizes[0x1] = dpb_size;
601 	buf_sizes[0x2] = image_size;
602 	buf_sizes[0x4] = min_ctx_size;
603 	return 0;
604 }
605 
606 /**
607  * amdgpu_uvd_cs_msg - handle UVD message
608  *
609  * @ctx: UVD parser context
610  * @bo: buffer object containing the message
611  * @offset: offset into the buffer object
612  *
613  * Peek into the UVD message and extract the session id.
614  * Make sure that we don't open up to many sessions.
615  */
616 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
617 			     struct amdgpu_bo *bo, unsigned offset)
618 {
619 	struct amdgpu_device *adev = ctx->parser->adev;
620 	int32_t *msg, msg_type, handle;
621 	void *ptr;
622 	long r;
623 	int i;
624 
625 	if (offset & 0x3F) {
626 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
627 		return -EINVAL;
628 	}
629 
630 	r = amdgpu_bo_kmap(bo, &ptr);
631 	if (r) {
632 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
633 		return r;
634 	}
635 
636 	msg = ptr + offset;
637 
638 	msg_type = msg[1];
639 	handle = msg[2];
640 
641 	if (handle == 0) {
642 		DRM_ERROR("Invalid UVD handle!\n");
643 		return -EINVAL;
644 	}
645 
646 	switch (msg_type) {
647 	case 0:
648 		/* it's a create msg, calc image size (width * height) */
649 		amdgpu_bo_kunmap(bo);
650 
651 		/* try to alloc a new handle */
652 		for (i = 0; i < adev->uvd.max_handles; ++i) {
653 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
654 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
655 				return -EINVAL;
656 			}
657 
658 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
659 				adev->uvd.filp[i] = ctx->parser->filp;
660 				return 0;
661 			}
662 		}
663 
664 		DRM_ERROR("No more free UVD handles!\n");
665 		return -EINVAL;
666 
667 	case 1:
668 		/* it's a decode msg, calc buffer sizes */
669 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
670 		amdgpu_bo_kunmap(bo);
671 		if (r)
672 			return r;
673 
674 		/* validate the handle */
675 		for (i = 0; i < adev->uvd.max_handles; ++i) {
676 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
677 				if (adev->uvd.filp[i] != ctx->parser->filp) {
678 					DRM_ERROR("UVD handle collision detected!\n");
679 					return -EINVAL;
680 				}
681 				return 0;
682 			}
683 		}
684 
685 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
686 		return -ENOENT;
687 
688 	case 2:
689 		/* it's a destroy msg, free the handle */
690 		for (i = 0; i < adev->uvd.max_handles; ++i)
691 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
692 		amdgpu_bo_kunmap(bo);
693 		return 0;
694 
695 	default:
696 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
697 		return -EINVAL;
698 	}
699 	BUG();
700 	return -EINVAL;
701 }
702 
703 /**
704  * amdgpu_uvd_cs_pass2 - second parsing round
705  *
706  * @ctx: UVD parser context
707  *
708  * Patch buffer addresses, make sure buffer sizes are correct.
709  */
710 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
711 {
712 	struct amdgpu_bo_va_mapping *mapping;
713 	struct amdgpu_bo *bo;
714 	uint32_t cmd, lo, hi;
715 	uint64_t start, end;
716 	uint64_t addr;
717 	int r;
718 
719 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
720 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
721 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
722 
723 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
724 	if (mapping == NULL)
725 		return -EINVAL;
726 
727 	start = amdgpu_bo_gpu_offset(bo);
728 
729 	end = (mapping->it.last + 1 - mapping->it.start);
730 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
731 
732 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
733 	start += addr;
734 
735 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
736 			    lower_32_bits(start));
737 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
738 			    upper_32_bits(start));
739 
740 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
741 	if (cmd < 0x4) {
742 		if ((end - start) < ctx->buf_sizes[cmd]) {
743 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
744 				  (unsigned)(end - start),
745 				  ctx->buf_sizes[cmd]);
746 			return -EINVAL;
747 		}
748 
749 	} else if (cmd == 0x206) {
750 		if ((end - start) < ctx->buf_sizes[4]) {
751 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
752 					  (unsigned)(end - start),
753 					  ctx->buf_sizes[4]);
754 			return -EINVAL;
755 		}
756 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
757 		DRM_ERROR("invalid UVD command %X!\n", cmd);
758 		return -EINVAL;
759 	}
760 
761 	if (!ctx->parser->adev->uvd.address_64_bit) {
762 		if ((start >> 28) != ((end - 1) >> 28)) {
763 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
764 				  start, end);
765 			return -EINVAL;
766 		}
767 
768 		if ((cmd == 0 || cmd == 0x3) &&
769 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
770 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
771 				  start, end);
772 			return -EINVAL;
773 		}
774 	}
775 
776 	if (cmd == 0) {
777 		ctx->has_msg_cmd = true;
778 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
779 		if (r)
780 			return r;
781 	} else if (!ctx->has_msg_cmd) {
782 		DRM_ERROR("Message needed before other commands are send!\n");
783 		return -EINVAL;
784 	}
785 
786 	return 0;
787 }
788 
789 /**
790  * amdgpu_uvd_cs_reg - parse register writes
791  *
792  * @ctx: UVD parser context
793  * @cb: callback function
794  *
795  * Parse the register writes, call cb on each complete command.
796  */
797 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
798 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
799 {
800 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
801 	int i, r;
802 
803 	ctx->idx++;
804 	for (i = 0; i <= ctx->count; ++i) {
805 		unsigned reg = ctx->reg + i;
806 
807 		if (ctx->idx >= ib->length_dw) {
808 			DRM_ERROR("Register command after end of CS!\n");
809 			return -EINVAL;
810 		}
811 
812 		switch (reg) {
813 		case mmUVD_GPCOM_VCPU_DATA0:
814 			ctx->data0 = ctx->idx;
815 			break;
816 		case mmUVD_GPCOM_VCPU_DATA1:
817 			ctx->data1 = ctx->idx;
818 			break;
819 		case mmUVD_GPCOM_VCPU_CMD:
820 			r = cb(ctx);
821 			if (r)
822 				return r;
823 			break;
824 		case mmUVD_ENGINE_CNTL:
825 			break;
826 		default:
827 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
828 			return -EINVAL;
829 		}
830 		ctx->idx++;
831 	}
832 	return 0;
833 }
834 
835 /**
836  * amdgpu_uvd_cs_packets - parse UVD packets
837  *
838  * @ctx: UVD parser context
839  * @cb: callback function
840  *
841  * Parse the command stream packets.
842  */
843 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
844 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
845 {
846 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
847 	int r;
848 
849 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
850 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
851 		unsigned type = CP_PACKET_GET_TYPE(cmd);
852 		switch (type) {
853 		case PACKET_TYPE0:
854 			ctx->reg = CP_PACKET0_GET_REG(cmd);
855 			ctx->count = CP_PACKET_GET_COUNT(cmd);
856 			r = amdgpu_uvd_cs_reg(ctx, cb);
857 			if (r)
858 				return r;
859 			break;
860 		case PACKET_TYPE2:
861 			++ctx->idx;
862 			break;
863 		default:
864 			DRM_ERROR("Unknown packet type %d !\n", type);
865 			return -EINVAL;
866 		}
867 	}
868 	return 0;
869 }
870 
871 /**
872  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
873  *
874  * @parser: Command submission parser context
875  *
876  * Parse the command stream, patch in addresses as necessary.
877  */
878 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
879 {
880 	struct amdgpu_uvd_cs_ctx ctx = {};
881 	unsigned buf_sizes[] = {
882 		[0x00000000]	=	2048,
883 		[0x00000001]	=	0xFFFFFFFF,
884 		[0x00000002]	=	0xFFFFFFFF,
885 		[0x00000003]	=	2048,
886 		[0x00000004]	=	0xFFFFFFFF,
887 	};
888 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
889 	int r;
890 
891 	if (ib->length_dw % 16) {
892 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
893 			  ib->length_dw);
894 		return -EINVAL;
895 	}
896 
897 	ctx.parser = parser;
898 	ctx.buf_sizes = buf_sizes;
899 	ctx.ib_idx = ib_idx;
900 
901 	/* first round, make sure the buffers are actually in the UVD segment */
902 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
903 	if (r)
904 		return r;
905 
906 	/* second round, patch buffer addresses into the command stream */
907 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
908 	if (r)
909 		return r;
910 
911 	if (!ctx.has_msg_cmd) {
912 		DRM_ERROR("UVD-IBs need a msg command!\n");
913 		return -EINVAL;
914 	}
915 
916 	amdgpu_uvd_note_usage(ctx.parser->adev);
917 
918 	return 0;
919 }
920 
921 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
922 			       bool direct, struct fence **fence)
923 {
924 	struct ttm_validate_buffer tv;
925 	struct ww_acquire_ctx ticket;
926 	struct list_head head;
927 	struct amdgpu_job *job;
928 	struct amdgpu_ib *ib;
929 	struct fence *f = NULL;
930 	struct amdgpu_device *adev = ring->adev;
931 	uint64_t addr;
932 	int i, r;
933 
934 	memset(&tv, 0, sizeof(tv));
935 	tv.bo = &bo->tbo;
936 
937 	INIT_LIST_HEAD(&head);
938 	list_add(&tv.head, &head);
939 
940 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
941 	if (r)
942 		return r;
943 
944 	if (!bo->adev->uvd.address_64_bit) {
945 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
946 		amdgpu_uvd_force_into_uvd_segment(bo);
947 	}
948 
949 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
950 	if (r)
951 		goto err;
952 
953 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
954 	if (r)
955 		goto err;
956 
957 	ib = &job->ibs[0];
958 	addr = amdgpu_bo_gpu_offset(bo);
959 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
960 	ib->ptr[1] = addr;
961 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
962 	ib->ptr[3] = addr >> 32;
963 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
964 	ib->ptr[5] = 0;
965 	for (i = 6; i < 16; ++i)
966 		ib->ptr[i] = PACKET2(0);
967 	ib->length_dw = 16;
968 
969 	if (direct) {
970 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, NULL, &f);
971 		job->fence = f;
972 		if (r)
973 			goto err_free;
974 
975 		amdgpu_job_free(job);
976 	} else {
977 		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
978 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
979 		if (r)
980 			goto err_free;
981 	}
982 
983 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
984 
985 	if (fence)
986 		*fence = fence_get(f);
987 	amdgpu_bo_unref(&bo);
988 	fence_put(f);
989 
990 	return 0;
991 
992 err_free:
993 	amdgpu_job_free(job);
994 
995 err:
996 	ttm_eu_backoff_reservation(&ticket, &head);
997 	return r;
998 }
999 
1000 /* multiple fence commands without any stream commands in between can
1001    crash the vcpu so just try to emmit a dummy create/destroy msg to
1002    avoid this */
1003 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1004 			      struct fence **fence)
1005 {
1006 	struct amdgpu_device *adev = ring->adev;
1007 	struct amdgpu_bo *bo;
1008 	uint32_t *msg;
1009 	int r, i;
1010 
1011 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1012 			     AMDGPU_GEM_DOMAIN_VRAM,
1013 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1014 			     NULL, NULL, &bo);
1015 	if (r)
1016 		return r;
1017 
1018 	r = amdgpu_bo_reserve(bo, false);
1019 	if (r) {
1020 		amdgpu_bo_unref(&bo);
1021 		return r;
1022 	}
1023 
1024 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1025 	if (r) {
1026 		amdgpu_bo_unreserve(bo);
1027 		amdgpu_bo_unref(&bo);
1028 		return r;
1029 	}
1030 
1031 	/* stitch together an UVD create msg */
1032 	msg[0] = cpu_to_le32(0x00000de4);
1033 	msg[1] = cpu_to_le32(0x00000000);
1034 	msg[2] = cpu_to_le32(handle);
1035 	msg[3] = cpu_to_le32(0x00000000);
1036 	msg[4] = cpu_to_le32(0x00000000);
1037 	msg[5] = cpu_to_le32(0x00000000);
1038 	msg[6] = cpu_to_le32(0x00000000);
1039 	msg[7] = cpu_to_le32(0x00000780);
1040 	msg[8] = cpu_to_le32(0x00000440);
1041 	msg[9] = cpu_to_le32(0x00000000);
1042 	msg[10] = cpu_to_le32(0x01b37000);
1043 	for (i = 11; i < 1024; ++i)
1044 		msg[i] = cpu_to_le32(0x0);
1045 
1046 	amdgpu_bo_kunmap(bo);
1047 	amdgpu_bo_unreserve(bo);
1048 
1049 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1050 }
1051 
1052 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1053 			       bool direct, struct fence **fence)
1054 {
1055 	struct amdgpu_device *adev = ring->adev;
1056 	struct amdgpu_bo *bo;
1057 	uint32_t *msg;
1058 	int r, i;
1059 
1060 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1061 			     AMDGPU_GEM_DOMAIN_VRAM,
1062 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1063 			     NULL, NULL, &bo);
1064 	if (r)
1065 		return r;
1066 
1067 	r = amdgpu_bo_reserve(bo, false);
1068 	if (r) {
1069 		amdgpu_bo_unref(&bo);
1070 		return r;
1071 	}
1072 
1073 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1074 	if (r) {
1075 		amdgpu_bo_unreserve(bo);
1076 		amdgpu_bo_unref(&bo);
1077 		return r;
1078 	}
1079 
1080 	/* stitch together an UVD destroy msg */
1081 	msg[0] = cpu_to_le32(0x00000de4);
1082 	msg[1] = cpu_to_le32(0x00000002);
1083 	msg[2] = cpu_to_le32(handle);
1084 	msg[3] = cpu_to_le32(0x00000000);
1085 	for (i = 4; i < 1024; ++i)
1086 		msg[i] = cpu_to_le32(0x0);
1087 
1088 	amdgpu_bo_kunmap(bo);
1089 	amdgpu_bo_unreserve(bo);
1090 
1091 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1092 }
1093 
1094 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1095 {
1096 	struct amdgpu_device *adev =
1097 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1098 	unsigned i, fences, handles = 0;
1099 
1100 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1101 
1102 	for (i = 0; i < adev->uvd.max_handles; ++i)
1103 		if (atomic_read(&adev->uvd.handles[i]))
1104 			++handles;
1105 
1106 	if (fences == 0 && handles == 0) {
1107 		if (adev->pm.dpm_enabled) {
1108 			amdgpu_dpm_enable_uvd(adev, false);
1109 		} else {
1110 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1111 		}
1112 	} else {
1113 		schedule_delayed_work(&adev->uvd.idle_work,
1114 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1115 	}
1116 }
1117 
1118 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1119 {
1120 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1121 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1122 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1123 
1124 	if (set_clocks) {
1125 		if (adev->pm.dpm_enabled) {
1126 			amdgpu_dpm_enable_uvd(adev, true);
1127 		} else {
1128 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1129 		}
1130 	}
1131 }
1132