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 	if (adev->uvd.vcpu_bo == NULL)
257 		return 0;
258 
259 	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
260 
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 	amdgpu_ring_fini(&adev->uvd.ring);
271 
272 	release_firmware(adev->uvd.fw);
273 
274 	return 0;
275 }
276 
277 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
278 {
279 	unsigned size;
280 	void *ptr;
281 	int i;
282 
283 	if (adev->uvd.vcpu_bo == NULL)
284 		return 0;
285 
286 	for (i = 0; i < adev->uvd.max_handles; ++i)
287 		if (atomic_read(&adev->uvd.handles[i]))
288 			break;
289 
290 	if (i == AMDGPU_MAX_UVD_HANDLES)
291 		return 0;
292 
293 	cancel_delayed_work_sync(&adev->uvd.idle_work);
294 
295 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
296 	ptr = adev->uvd.cpu_addr;
297 
298 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
299 	if (!adev->uvd.saved_bo)
300 		return -ENOMEM;
301 
302 	memcpy(adev->uvd.saved_bo, ptr, size);
303 
304 	return 0;
305 }
306 
307 int amdgpu_uvd_resume(struct amdgpu_device *adev)
308 {
309 	unsigned size;
310 	void *ptr;
311 
312 	if (adev->uvd.vcpu_bo == NULL)
313 		return -EINVAL;
314 
315 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
316 	ptr = adev->uvd.cpu_addr;
317 
318 	if (adev->uvd.saved_bo != NULL) {
319 		memcpy(ptr, adev->uvd.saved_bo, size);
320 		kfree(adev->uvd.saved_bo);
321 		adev->uvd.saved_bo = NULL;
322 	} else {
323 		const struct common_firmware_header *hdr;
324 		unsigned offset;
325 
326 		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
327 		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
328 		memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
329 			(adev->uvd.fw->size) - offset);
330 		size -= le32_to_cpu(hdr->ucode_size_bytes);
331 		ptr += le32_to_cpu(hdr->ucode_size_bytes);
332 		memset(ptr, 0, size);
333 	}
334 
335 	return 0;
336 }
337 
338 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
339 {
340 	struct amdgpu_ring *ring = &adev->uvd.ring;
341 	int i, r;
342 
343 	for (i = 0; i < adev->uvd.max_handles; ++i) {
344 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
345 		if (handle != 0 && adev->uvd.filp[i] == filp) {
346 			struct fence *fence;
347 
348 			amdgpu_uvd_note_usage(adev);
349 
350 			r = amdgpu_uvd_get_destroy_msg(ring, handle,
351 						       false, &fence);
352 			if (r) {
353 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
354 				continue;
355 			}
356 
357 			fence_wait(fence, false);
358 			fence_put(fence);
359 
360 			adev->uvd.filp[i] = NULL;
361 			atomic_set(&adev->uvd.handles[i], 0);
362 		}
363 	}
364 }
365 
366 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
367 {
368 	int i;
369 	for (i = 0; i < rbo->placement.num_placement; ++i) {
370 		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
371 		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
372 	}
373 }
374 
375 /**
376  * amdgpu_uvd_cs_pass1 - first parsing round
377  *
378  * @ctx: UVD parser context
379  *
380  * Make sure UVD message and feedback buffers are in VRAM and
381  * nobody is violating an 256MB boundary.
382  */
383 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
384 {
385 	struct amdgpu_bo_va_mapping *mapping;
386 	struct amdgpu_bo *bo;
387 	uint32_t cmd, lo, hi;
388 	uint64_t addr;
389 	int r = 0;
390 
391 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
392 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
393 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
394 
395 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
396 	if (mapping == NULL) {
397 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
398 		return -EINVAL;
399 	}
400 
401 	if (!ctx->parser->adev->uvd.address_64_bit) {
402 		/* check if it's a message or feedback command */
403 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
404 		if (cmd == 0x0 || cmd == 0x3) {
405 			/* yes, force it into VRAM */
406 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
407 			amdgpu_ttm_placement_from_domain(bo, domain);
408 		}
409 		amdgpu_uvd_force_into_uvd_segment(bo);
410 
411 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
412 	}
413 
414 	return r;
415 }
416 
417 /**
418  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
419  *
420  * @msg: pointer to message structure
421  * @buf_sizes: returned buffer sizes
422  *
423  * Peek into the decode message and calculate the necessary buffer sizes.
424  */
425 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
426 	unsigned buf_sizes[])
427 {
428 	unsigned stream_type = msg[4];
429 	unsigned width = msg[6];
430 	unsigned height = msg[7];
431 	unsigned dpb_size = msg[9];
432 	unsigned pitch = msg[28];
433 	unsigned level = msg[57];
434 
435 	unsigned width_in_mb = width / 16;
436 	unsigned height_in_mb = ALIGN(height / 16, 2);
437 	unsigned fs_in_mb = width_in_mb * height_in_mb;
438 
439 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
440 	unsigned min_ctx_size = 0;
441 
442 	image_size = width * height;
443 	image_size += image_size / 2;
444 	image_size = ALIGN(image_size, 1024);
445 
446 	switch (stream_type) {
447 	case 0: /* H264 */
448 		switch(level) {
449 		case 30:
450 			num_dpb_buffer = 8100 / fs_in_mb;
451 			break;
452 		case 31:
453 			num_dpb_buffer = 18000 / fs_in_mb;
454 			break;
455 		case 32:
456 			num_dpb_buffer = 20480 / fs_in_mb;
457 			break;
458 		case 41:
459 			num_dpb_buffer = 32768 / fs_in_mb;
460 			break;
461 		case 42:
462 			num_dpb_buffer = 34816 / fs_in_mb;
463 			break;
464 		case 50:
465 			num_dpb_buffer = 110400 / fs_in_mb;
466 			break;
467 		case 51:
468 			num_dpb_buffer = 184320 / fs_in_mb;
469 			break;
470 		default:
471 			num_dpb_buffer = 184320 / fs_in_mb;
472 			break;
473 		}
474 		num_dpb_buffer++;
475 		if (num_dpb_buffer > 17)
476 			num_dpb_buffer = 17;
477 
478 		/* reference picture buffer */
479 		min_dpb_size = image_size * num_dpb_buffer;
480 
481 		/* macroblock context buffer */
482 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
483 
484 		/* IT surface buffer */
485 		min_dpb_size += width_in_mb * height_in_mb * 32;
486 		break;
487 
488 	case 1: /* VC1 */
489 
490 		/* reference picture buffer */
491 		min_dpb_size = image_size * 3;
492 
493 		/* CONTEXT_BUFFER */
494 		min_dpb_size += width_in_mb * height_in_mb * 128;
495 
496 		/* IT surface buffer */
497 		min_dpb_size += width_in_mb * 64;
498 
499 		/* DB surface buffer */
500 		min_dpb_size += width_in_mb * 128;
501 
502 		/* BP */
503 		tmp = max(width_in_mb, height_in_mb);
504 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
505 		break;
506 
507 	case 3: /* MPEG2 */
508 
509 		/* reference picture buffer */
510 		min_dpb_size = image_size * 3;
511 		break;
512 
513 	case 4: /* MPEG4 */
514 
515 		/* reference picture buffer */
516 		min_dpb_size = image_size * 3;
517 
518 		/* CM */
519 		min_dpb_size += width_in_mb * height_in_mb * 64;
520 
521 		/* IT surface buffer */
522 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
523 		break;
524 
525 	case 7: /* H264 Perf */
526 		switch(level) {
527 		case 30:
528 			num_dpb_buffer = 8100 / fs_in_mb;
529 			break;
530 		case 31:
531 			num_dpb_buffer = 18000 / fs_in_mb;
532 			break;
533 		case 32:
534 			num_dpb_buffer = 20480 / fs_in_mb;
535 			break;
536 		case 41:
537 			num_dpb_buffer = 32768 / fs_in_mb;
538 			break;
539 		case 42:
540 			num_dpb_buffer = 34816 / fs_in_mb;
541 			break;
542 		case 50:
543 			num_dpb_buffer = 110400 / fs_in_mb;
544 			break;
545 		case 51:
546 			num_dpb_buffer = 184320 / fs_in_mb;
547 			break;
548 		default:
549 			num_dpb_buffer = 184320 / fs_in_mb;
550 			break;
551 		}
552 		num_dpb_buffer++;
553 		if (num_dpb_buffer > 17)
554 			num_dpb_buffer = 17;
555 
556 		/* reference picture buffer */
557 		min_dpb_size = image_size * num_dpb_buffer;
558 
559 		if (adev->asic_type < CHIP_POLARIS10){
560 			/* macroblock context buffer */
561 			min_dpb_size +=
562 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
563 
564 			/* IT surface buffer */
565 			min_dpb_size += width_in_mb * height_in_mb * 32;
566 		} else {
567 			/* macroblock context buffer */
568 			min_ctx_size =
569 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
570 		}
571 		break;
572 
573 	case 16: /* H265 */
574 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
575 		image_size = ALIGN(image_size, 256);
576 
577 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
578 		min_dpb_size = image_size * num_dpb_buffer;
579 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
580 					   * 16 * num_dpb_buffer + 52 * 1024;
581 		break;
582 
583 	default:
584 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
585 		return -EINVAL;
586 	}
587 
588 	if (width > pitch) {
589 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
590 		return -EINVAL;
591 	}
592 
593 	if (dpb_size < min_dpb_size) {
594 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
595 			  dpb_size, min_dpb_size);
596 		return -EINVAL;
597 	}
598 
599 	buf_sizes[0x1] = dpb_size;
600 	buf_sizes[0x2] = image_size;
601 	buf_sizes[0x4] = min_ctx_size;
602 	return 0;
603 }
604 
605 /**
606  * amdgpu_uvd_cs_msg - handle UVD message
607  *
608  * @ctx: UVD parser context
609  * @bo: buffer object containing the message
610  * @offset: offset into the buffer object
611  *
612  * Peek into the UVD message and extract the session id.
613  * Make sure that we don't open up to many sessions.
614  */
615 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
616 			     struct amdgpu_bo *bo, unsigned offset)
617 {
618 	struct amdgpu_device *adev = ctx->parser->adev;
619 	int32_t *msg, msg_type, handle;
620 	void *ptr;
621 	long r;
622 	int i;
623 
624 	if (offset & 0x3F) {
625 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
626 		return -EINVAL;
627 	}
628 
629 	r = amdgpu_bo_kmap(bo, &ptr);
630 	if (r) {
631 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
632 		return r;
633 	}
634 
635 	msg = ptr + offset;
636 
637 	msg_type = msg[1];
638 	handle = msg[2];
639 
640 	if (handle == 0) {
641 		DRM_ERROR("Invalid UVD handle!\n");
642 		return -EINVAL;
643 	}
644 
645 	switch (msg_type) {
646 	case 0:
647 		/* it's a create msg, calc image size (width * height) */
648 		amdgpu_bo_kunmap(bo);
649 
650 		/* try to alloc a new handle */
651 		for (i = 0; i < adev->uvd.max_handles; ++i) {
652 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
653 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
654 				return -EINVAL;
655 			}
656 
657 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
658 				adev->uvd.filp[i] = ctx->parser->filp;
659 				return 0;
660 			}
661 		}
662 
663 		DRM_ERROR("No more free UVD handles!\n");
664 		return -EINVAL;
665 
666 	case 1:
667 		/* it's a decode msg, calc buffer sizes */
668 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
669 		amdgpu_bo_kunmap(bo);
670 		if (r)
671 			return r;
672 
673 		/* validate the handle */
674 		for (i = 0; i < adev->uvd.max_handles; ++i) {
675 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
676 				if (adev->uvd.filp[i] != ctx->parser->filp) {
677 					DRM_ERROR("UVD handle collision detected!\n");
678 					return -EINVAL;
679 				}
680 				return 0;
681 			}
682 		}
683 
684 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
685 		return -ENOENT;
686 
687 	case 2:
688 		/* it's a destroy msg, free the handle */
689 		for (i = 0; i < adev->uvd.max_handles; ++i)
690 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
691 		amdgpu_bo_kunmap(bo);
692 		return 0;
693 
694 	default:
695 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
696 		return -EINVAL;
697 	}
698 	BUG();
699 	return -EINVAL;
700 }
701 
702 /**
703  * amdgpu_uvd_cs_pass2 - second parsing round
704  *
705  * @ctx: UVD parser context
706  *
707  * Patch buffer addresses, make sure buffer sizes are correct.
708  */
709 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
710 {
711 	struct amdgpu_bo_va_mapping *mapping;
712 	struct amdgpu_bo *bo;
713 	uint32_t cmd, lo, hi;
714 	uint64_t start, end;
715 	uint64_t addr;
716 	int r;
717 
718 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
719 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
720 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
721 
722 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
723 	if (mapping == NULL)
724 		return -EINVAL;
725 
726 	start = amdgpu_bo_gpu_offset(bo);
727 
728 	end = (mapping->it.last + 1 - mapping->it.start);
729 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
730 
731 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
732 	start += addr;
733 
734 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
735 			    lower_32_bits(start));
736 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
737 			    upper_32_bits(start));
738 
739 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
740 	if (cmd < 0x4) {
741 		if ((end - start) < ctx->buf_sizes[cmd]) {
742 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
743 				  (unsigned)(end - start),
744 				  ctx->buf_sizes[cmd]);
745 			return -EINVAL;
746 		}
747 
748 	} else if (cmd == 0x206) {
749 		if ((end - start) < ctx->buf_sizes[4]) {
750 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
751 					  (unsigned)(end - start),
752 					  ctx->buf_sizes[4]);
753 			return -EINVAL;
754 		}
755 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
756 		DRM_ERROR("invalid UVD command %X!\n", cmd);
757 		return -EINVAL;
758 	}
759 
760 	if (!ctx->parser->adev->uvd.address_64_bit) {
761 		if ((start >> 28) != ((end - 1) >> 28)) {
762 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
763 				  start, end);
764 			return -EINVAL;
765 		}
766 
767 		if ((cmd == 0 || cmd == 0x3) &&
768 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
769 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
770 				  start, end);
771 			return -EINVAL;
772 		}
773 	}
774 
775 	if (cmd == 0) {
776 		ctx->has_msg_cmd = true;
777 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
778 		if (r)
779 			return r;
780 	} else if (!ctx->has_msg_cmd) {
781 		DRM_ERROR("Message needed before other commands are send!\n");
782 		return -EINVAL;
783 	}
784 
785 	return 0;
786 }
787 
788 /**
789  * amdgpu_uvd_cs_reg - parse register writes
790  *
791  * @ctx: UVD parser context
792  * @cb: callback function
793  *
794  * Parse the register writes, call cb on each complete command.
795  */
796 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
797 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
798 {
799 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
800 	int i, r;
801 
802 	ctx->idx++;
803 	for (i = 0; i <= ctx->count; ++i) {
804 		unsigned reg = ctx->reg + i;
805 
806 		if (ctx->idx >= ib->length_dw) {
807 			DRM_ERROR("Register command after end of CS!\n");
808 			return -EINVAL;
809 		}
810 
811 		switch (reg) {
812 		case mmUVD_GPCOM_VCPU_DATA0:
813 			ctx->data0 = ctx->idx;
814 			break;
815 		case mmUVD_GPCOM_VCPU_DATA1:
816 			ctx->data1 = ctx->idx;
817 			break;
818 		case mmUVD_GPCOM_VCPU_CMD:
819 			r = cb(ctx);
820 			if (r)
821 				return r;
822 			break;
823 		case mmUVD_ENGINE_CNTL:
824 			break;
825 		default:
826 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
827 			return -EINVAL;
828 		}
829 		ctx->idx++;
830 	}
831 	return 0;
832 }
833 
834 /**
835  * amdgpu_uvd_cs_packets - parse UVD packets
836  *
837  * @ctx: UVD parser context
838  * @cb: callback function
839  *
840  * Parse the command stream packets.
841  */
842 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
843 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
844 {
845 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
846 	int r;
847 
848 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
849 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
850 		unsigned type = CP_PACKET_GET_TYPE(cmd);
851 		switch (type) {
852 		case PACKET_TYPE0:
853 			ctx->reg = CP_PACKET0_GET_REG(cmd);
854 			ctx->count = CP_PACKET_GET_COUNT(cmd);
855 			r = amdgpu_uvd_cs_reg(ctx, cb);
856 			if (r)
857 				return r;
858 			break;
859 		case PACKET_TYPE2:
860 			++ctx->idx;
861 			break;
862 		default:
863 			DRM_ERROR("Unknown packet type %d !\n", type);
864 			return -EINVAL;
865 		}
866 	}
867 	return 0;
868 }
869 
870 /**
871  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
872  *
873  * @parser: Command submission parser context
874  *
875  * Parse the command stream, patch in addresses as necessary.
876  */
877 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
878 {
879 	struct amdgpu_uvd_cs_ctx ctx = {};
880 	unsigned buf_sizes[] = {
881 		[0x00000000]	=	2048,
882 		[0x00000001]	=	0xFFFFFFFF,
883 		[0x00000002]	=	0xFFFFFFFF,
884 		[0x00000003]	=	2048,
885 		[0x00000004]	=	0xFFFFFFFF,
886 	};
887 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
888 	int r;
889 
890 	if (ib->length_dw % 16) {
891 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
892 			  ib->length_dw);
893 		return -EINVAL;
894 	}
895 
896 	ctx.parser = parser;
897 	ctx.buf_sizes = buf_sizes;
898 	ctx.ib_idx = ib_idx;
899 
900 	/* first round, make sure the buffers are actually in the UVD segment */
901 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
902 	if (r)
903 		return r;
904 
905 	/* second round, patch buffer addresses into the command stream */
906 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
907 	if (r)
908 		return r;
909 
910 	if (!ctx.has_msg_cmd) {
911 		DRM_ERROR("UVD-IBs need a msg command!\n");
912 		return -EINVAL;
913 	}
914 
915 	amdgpu_uvd_note_usage(ctx.parser->adev);
916 
917 	return 0;
918 }
919 
920 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
921 			       bool direct, struct fence **fence)
922 {
923 	struct ttm_validate_buffer tv;
924 	struct ww_acquire_ctx ticket;
925 	struct list_head head;
926 	struct amdgpu_job *job;
927 	struct amdgpu_ib *ib;
928 	struct fence *f = NULL;
929 	struct amdgpu_device *adev = ring->adev;
930 	uint64_t addr;
931 	int i, r;
932 
933 	memset(&tv, 0, sizeof(tv));
934 	tv.bo = &bo->tbo;
935 
936 	INIT_LIST_HEAD(&head);
937 	list_add(&tv.head, &head);
938 
939 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
940 	if (r)
941 		return r;
942 
943 	if (!bo->adev->uvd.address_64_bit) {
944 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
945 		amdgpu_uvd_force_into_uvd_segment(bo);
946 	}
947 
948 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
949 	if (r)
950 		goto err;
951 
952 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
953 	if (r)
954 		goto err;
955 
956 	ib = &job->ibs[0];
957 	addr = amdgpu_bo_gpu_offset(bo);
958 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
959 	ib->ptr[1] = addr;
960 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
961 	ib->ptr[3] = addr >> 32;
962 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
963 	ib->ptr[5] = 0;
964 	for (i = 6; i < 16; ++i)
965 		ib->ptr[i] = PACKET2(0);
966 	ib->length_dw = 16;
967 
968 	if (direct) {
969 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, NULL, &f);
970 		job->fence = f;
971 		if (r)
972 			goto err_free;
973 
974 		amdgpu_job_free(job);
975 	} else {
976 		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
977 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
978 		if (r)
979 			goto err_free;
980 	}
981 
982 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
983 
984 	if (fence)
985 		*fence = fence_get(f);
986 	amdgpu_bo_unref(&bo);
987 	fence_put(f);
988 
989 	return 0;
990 
991 err_free:
992 	amdgpu_job_free(job);
993 
994 err:
995 	ttm_eu_backoff_reservation(&ticket, &head);
996 	return r;
997 }
998 
999 /* multiple fence commands without any stream commands in between can
1000    crash the vcpu so just try to emmit a dummy create/destroy msg to
1001    avoid this */
1002 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1003 			      struct fence **fence)
1004 {
1005 	struct amdgpu_device *adev = ring->adev;
1006 	struct amdgpu_bo *bo;
1007 	uint32_t *msg;
1008 	int r, i;
1009 
1010 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1011 			     AMDGPU_GEM_DOMAIN_VRAM,
1012 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1013 			     NULL, NULL, &bo);
1014 	if (r)
1015 		return r;
1016 
1017 	r = amdgpu_bo_reserve(bo, false);
1018 	if (r) {
1019 		amdgpu_bo_unref(&bo);
1020 		return r;
1021 	}
1022 
1023 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1024 	if (r) {
1025 		amdgpu_bo_unreserve(bo);
1026 		amdgpu_bo_unref(&bo);
1027 		return r;
1028 	}
1029 
1030 	/* stitch together an UVD create msg */
1031 	msg[0] = cpu_to_le32(0x00000de4);
1032 	msg[1] = cpu_to_le32(0x00000000);
1033 	msg[2] = cpu_to_le32(handle);
1034 	msg[3] = cpu_to_le32(0x00000000);
1035 	msg[4] = cpu_to_le32(0x00000000);
1036 	msg[5] = cpu_to_le32(0x00000000);
1037 	msg[6] = cpu_to_le32(0x00000000);
1038 	msg[7] = cpu_to_le32(0x00000780);
1039 	msg[8] = cpu_to_le32(0x00000440);
1040 	msg[9] = cpu_to_le32(0x00000000);
1041 	msg[10] = cpu_to_le32(0x01b37000);
1042 	for (i = 11; i < 1024; ++i)
1043 		msg[i] = cpu_to_le32(0x0);
1044 
1045 	amdgpu_bo_kunmap(bo);
1046 	amdgpu_bo_unreserve(bo);
1047 
1048 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1049 }
1050 
1051 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1052 			       bool direct, struct fence **fence)
1053 {
1054 	struct amdgpu_device *adev = ring->adev;
1055 	struct amdgpu_bo *bo;
1056 	uint32_t *msg;
1057 	int r, i;
1058 
1059 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1060 			     AMDGPU_GEM_DOMAIN_VRAM,
1061 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1062 			     NULL, NULL, &bo);
1063 	if (r)
1064 		return r;
1065 
1066 	r = amdgpu_bo_reserve(bo, false);
1067 	if (r) {
1068 		amdgpu_bo_unref(&bo);
1069 		return r;
1070 	}
1071 
1072 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1073 	if (r) {
1074 		amdgpu_bo_unreserve(bo);
1075 		amdgpu_bo_unref(&bo);
1076 		return r;
1077 	}
1078 
1079 	/* stitch together an UVD destroy msg */
1080 	msg[0] = cpu_to_le32(0x00000de4);
1081 	msg[1] = cpu_to_le32(0x00000002);
1082 	msg[2] = cpu_to_le32(handle);
1083 	msg[3] = cpu_to_le32(0x00000000);
1084 	for (i = 4; i < 1024; ++i)
1085 		msg[i] = cpu_to_le32(0x0);
1086 
1087 	amdgpu_bo_kunmap(bo);
1088 	amdgpu_bo_unreserve(bo);
1089 
1090 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1091 }
1092 
1093 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1094 {
1095 	struct amdgpu_device *adev =
1096 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1097 	unsigned i, fences, handles = 0;
1098 
1099 	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1100 
1101 	for (i = 0; i < adev->uvd.max_handles; ++i)
1102 		if (atomic_read(&adev->uvd.handles[i]))
1103 			++handles;
1104 
1105 	if (fences == 0 && handles == 0) {
1106 		if (adev->pm.dpm_enabled) {
1107 			amdgpu_dpm_enable_uvd(adev, false);
1108 		} else {
1109 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1110 		}
1111 	} else {
1112 		schedule_delayed_work(&adev->uvd.idle_work,
1113 				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1114 	}
1115 }
1116 
1117 static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1118 {
1119 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1120 	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1121 					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1122 
1123 	if (set_clocks) {
1124 		if (adev->pm.dpm_enabled) {
1125 			amdgpu_dpm_enable_uvd(adev, true);
1126 		} else {
1127 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1128 		}
1129 	}
1130 }
1131