xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
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	msecs_to_jiffies(1000)
44 
45 /* Firmware versions for VI */
46 #define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
47 #define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
48 #define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
49 #define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
50 
51 /* Polaris10/11 firmware version */
52 #define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
53 
54 /* Firmware Names */
55 #ifdef CONFIG_DRM_AMDGPU_CIK
56 #define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
57 #define FIRMWARE_KABINI	"radeon/kabini_uvd.bin"
58 #define FIRMWARE_KAVERI	"radeon/kaveri_uvd.bin"
59 #define FIRMWARE_HAWAII	"radeon/hawaii_uvd.bin"
60 #define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
61 #endif
62 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
63 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
64 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
65 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
66 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
67 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
68 
69 /**
70  * amdgpu_uvd_cs_ctx - Command submission parser context
71  *
72  * Used for emulating virtual memory support on UVD 4.2.
73  */
74 struct amdgpu_uvd_cs_ctx {
75 	struct amdgpu_cs_parser *parser;
76 	unsigned reg, count;
77 	unsigned data0, data1;
78 	unsigned idx;
79 	unsigned ib_idx;
80 
81 	/* does the IB has a msg command */
82 	bool has_msg_cmd;
83 
84 	/* minimum buffer sizes */
85 	unsigned *buf_sizes;
86 };
87 
88 #ifdef CONFIG_DRM_AMDGPU_CIK
89 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
90 MODULE_FIRMWARE(FIRMWARE_KABINI);
91 MODULE_FIRMWARE(FIRMWARE_KAVERI);
92 MODULE_FIRMWARE(FIRMWARE_HAWAII);
93 MODULE_FIRMWARE(FIRMWARE_MULLINS);
94 #endif
95 MODULE_FIRMWARE(FIRMWARE_TONGA);
96 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
97 MODULE_FIRMWARE(FIRMWARE_FIJI);
98 MODULE_FIRMWARE(FIRMWARE_STONEY);
99 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
100 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
101 
102 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
103 
104 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
105 {
106 	struct amdgpu_ring *ring;
107 	struct amd_sched_rq *rq;
108 	unsigned long bo_size;
109 	const char *fw_name;
110 	const struct common_firmware_header *hdr;
111 	unsigned version_major, version_minor, family_id;
112 	int i, r;
113 
114 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
115 
116 	switch (adev->asic_type) {
117 #ifdef CONFIG_DRM_AMDGPU_CIK
118 	case CHIP_BONAIRE:
119 		fw_name = FIRMWARE_BONAIRE;
120 		break;
121 	case CHIP_KABINI:
122 		fw_name = FIRMWARE_KABINI;
123 		break;
124 	case CHIP_KAVERI:
125 		fw_name = FIRMWARE_KAVERI;
126 		break;
127 	case CHIP_HAWAII:
128 		fw_name = FIRMWARE_HAWAII;
129 		break;
130 	case CHIP_MULLINS:
131 		fw_name = FIRMWARE_MULLINS;
132 		break;
133 #endif
134 	case CHIP_TONGA:
135 		fw_name = FIRMWARE_TONGA;
136 		break;
137 	case CHIP_FIJI:
138 		fw_name = FIRMWARE_FIJI;
139 		break;
140 	case CHIP_CARRIZO:
141 		fw_name = FIRMWARE_CARRIZO;
142 		break;
143 	case CHIP_STONEY:
144 		fw_name = FIRMWARE_STONEY;
145 		break;
146 	case CHIP_POLARIS10:
147 		fw_name = FIRMWARE_POLARIS10;
148 		break;
149 	case CHIP_POLARIS11:
150 		fw_name = FIRMWARE_POLARIS11;
151 		break;
152 	default:
153 		return -EINVAL;
154 	}
155 
156 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
157 	if (r) {
158 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
159 			fw_name);
160 		return r;
161 	}
162 
163 	r = amdgpu_ucode_validate(adev->uvd.fw);
164 	if (r) {
165 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
166 			fw_name);
167 		release_firmware(adev->uvd.fw);
168 		adev->uvd.fw = NULL;
169 		return r;
170 	}
171 
172 	/* Set the default UVD handles that the firmware can handle */
173 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
174 
175 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
176 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
177 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
178 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
179 	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
180 		version_major, version_minor, family_id);
181 
182 	/*
183 	 * Limit the number of UVD handles depending on microcode major
184 	 * and minor versions. The firmware version which has 40 UVD
185 	 * instances support is 1.80. So all subsequent versions should
186 	 * also have the same support.
187 	 */
188 	if ((version_major > 0x01) ||
189 	    ((version_major == 0x01) && (version_minor >= 0x50)))
190 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
191 
192 	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
193 				(family_id << 8));
194 
195 	if ((adev->asic_type == CHIP_POLARIS10 ||
196 	     adev->asic_type == CHIP_POLARIS11) &&
197 	    (adev->uvd.fw_version < FW_1_66_16))
198 		DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
199 			  version_major, version_minor);
200 
201 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
202 		  +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
203 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
204 	r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
205 				    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.vcpu_bo,
206 				    &adev->uvd.gpu_addr, &adev->uvd.cpu_addr);
207 	if (r) {
208 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
209 		return r;
210 	}
211 
212 	ring = &adev->uvd.ring;
213 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
214 	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
215 				  rq, amdgpu_sched_jobs);
216 	if (r != 0) {
217 		DRM_ERROR("Failed setting up UVD run queue.\n");
218 		return r;
219 	}
220 
221 	for (i = 0; i < adev->uvd.max_handles; ++i) {
222 		atomic_set(&adev->uvd.handles[i], 0);
223 		adev->uvd.filp[i] = NULL;
224 	}
225 
226 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
227 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
228 		adev->uvd.address_64_bit = true;
229 
230 	switch (adev->asic_type) {
231 	case CHIP_TONGA:
232 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
233 		break;
234 	case CHIP_CARRIZO:
235 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
236 		break;
237 	case CHIP_FIJI:
238 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
239 		break;
240 	case CHIP_STONEY:
241 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
242 		break;
243 	default:
244 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
245 	}
246 
247 	return 0;
248 }
249 
250 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
251 {
252 	kfree(adev->uvd.saved_bo);
253 
254 	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
255 
256 	amdgpu_bo_free_kernel(&adev->uvd.vcpu_bo,
257 			      &adev->uvd.gpu_addr,
258 			      (void **)&adev->uvd.cpu_addr);
259 
260 	amdgpu_ring_fini(&adev->uvd.ring);
261 
262 	release_firmware(adev->uvd.fw);
263 
264 	return 0;
265 }
266 
267 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
268 {
269 	unsigned size;
270 	void *ptr;
271 	int i;
272 
273 	if (adev->uvd.vcpu_bo == NULL)
274 		return 0;
275 
276 	for (i = 0; i < adev->uvd.max_handles; ++i)
277 		if (atomic_read(&adev->uvd.handles[i]))
278 			break;
279 
280 	if (i == AMDGPU_MAX_UVD_HANDLES)
281 		return 0;
282 
283 	cancel_delayed_work_sync(&adev->uvd.idle_work);
284 
285 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
286 	ptr = adev->uvd.cpu_addr;
287 
288 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
289 	if (!adev->uvd.saved_bo)
290 		return -ENOMEM;
291 
292 	memcpy_fromio(adev->uvd.saved_bo, ptr, size);
293 
294 	return 0;
295 }
296 
297 int amdgpu_uvd_resume(struct amdgpu_device *adev)
298 {
299 	unsigned size;
300 	void *ptr;
301 
302 	if (adev->uvd.vcpu_bo == NULL)
303 		return -EINVAL;
304 
305 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
306 	ptr = adev->uvd.cpu_addr;
307 
308 	if (adev->uvd.saved_bo != NULL) {
309 		memcpy_toio(ptr, adev->uvd.saved_bo, size);
310 		kfree(adev->uvd.saved_bo);
311 		adev->uvd.saved_bo = NULL;
312 	} else {
313 		const struct common_firmware_header *hdr;
314 		unsigned offset;
315 
316 		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
317 		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
318 		memcpy_toio(adev->uvd.cpu_addr, adev->uvd.fw->data + offset,
319 			    le32_to_cpu(hdr->ucode_size_bytes));
320 		size -= le32_to_cpu(hdr->ucode_size_bytes);
321 		ptr += le32_to_cpu(hdr->ucode_size_bytes);
322 		memset_io(ptr, 0, size);
323 	}
324 
325 	return 0;
326 }
327 
328 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
329 {
330 	struct amdgpu_ring *ring = &adev->uvd.ring;
331 	int i, r;
332 
333 	for (i = 0; i < adev->uvd.max_handles; ++i) {
334 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
335 		if (handle != 0 && adev->uvd.filp[i] == filp) {
336 			struct dma_fence *fence;
337 
338 			r = amdgpu_uvd_get_destroy_msg(ring, handle,
339 						       false, &fence);
340 			if (r) {
341 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
342 				continue;
343 			}
344 
345 			dma_fence_wait(fence, false);
346 			dma_fence_put(fence);
347 
348 			adev->uvd.filp[i] = NULL;
349 			atomic_set(&adev->uvd.handles[i], 0);
350 		}
351 	}
352 }
353 
354 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
355 {
356 	int i;
357 	for (i = 0; i < abo->placement.num_placement; ++i) {
358 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
359 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
360 	}
361 }
362 
363 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
364 {
365 	uint32_t lo, hi;
366 	uint64_t addr;
367 
368 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
369 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
370 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
371 
372 	return addr;
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;
388 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
389 	int r = 0;
390 
391 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
392 	if (mapping == NULL) {
393 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
394 		return -EINVAL;
395 	}
396 
397 	if (!ctx->parser->adev->uvd.address_64_bit) {
398 		/* check if it's a message or feedback command */
399 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
400 		if (cmd == 0x0 || cmd == 0x3) {
401 			/* yes, force it into VRAM */
402 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
403 			amdgpu_ttm_placement_from_domain(bo, domain);
404 		}
405 		amdgpu_uvd_force_into_uvd_segment(bo);
406 
407 		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
408 	}
409 
410 	return r;
411 }
412 
413 /**
414  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
415  *
416  * @msg: pointer to message structure
417  * @buf_sizes: returned buffer sizes
418  *
419  * Peek into the decode message and calculate the necessary buffer sizes.
420  */
421 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
422 	unsigned buf_sizes[])
423 {
424 	unsigned stream_type = msg[4];
425 	unsigned width = msg[6];
426 	unsigned height = msg[7];
427 	unsigned dpb_size = msg[9];
428 	unsigned pitch = msg[28];
429 	unsigned level = msg[57];
430 
431 	unsigned width_in_mb = width / 16;
432 	unsigned height_in_mb = ALIGN(height / 16, 2);
433 	unsigned fs_in_mb = width_in_mb * height_in_mb;
434 
435 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
436 	unsigned min_ctx_size = ~0;
437 
438 	image_size = width * height;
439 	image_size += image_size / 2;
440 	image_size = ALIGN(image_size, 1024);
441 
442 	switch (stream_type) {
443 	case 0: /* H264 */
444 		switch(level) {
445 		case 30:
446 			num_dpb_buffer = 8100 / fs_in_mb;
447 			break;
448 		case 31:
449 			num_dpb_buffer = 18000 / fs_in_mb;
450 			break;
451 		case 32:
452 			num_dpb_buffer = 20480 / fs_in_mb;
453 			break;
454 		case 41:
455 			num_dpb_buffer = 32768 / fs_in_mb;
456 			break;
457 		case 42:
458 			num_dpb_buffer = 34816 / fs_in_mb;
459 			break;
460 		case 50:
461 			num_dpb_buffer = 110400 / fs_in_mb;
462 			break;
463 		case 51:
464 			num_dpb_buffer = 184320 / fs_in_mb;
465 			break;
466 		default:
467 			num_dpb_buffer = 184320 / fs_in_mb;
468 			break;
469 		}
470 		num_dpb_buffer++;
471 		if (num_dpb_buffer > 17)
472 			num_dpb_buffer = 17;
473 
474 		/* reference picture buffer */
475 		min_dpb_size = image_size * num_dpb_buffer;
476 
477 		/* macroblock context buffer */
478 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
479 
480 		/* IT surface buffer */
481 		min_dpb_size += width_in_mb * height_in_mb * 32;
482 		break;
483 
484 	case 1: /* VC1 */
485 
486 		/* reference picture buffer */
487 		min_dpb_size = image_size * 3;
488 
489 		/* CONTEXT_BUFFER */
490 		min_dpb_size += width_in_mb * height_in_mb * 128;
491 
492 		/* IT surface buffer */
493 		min_dpb_size += width_in_mb * 64;
494 
495 		/* DB surface buffer */
496 		min_dpb_size += width_in_mb * 128;
497 
498 		/* BP */
499 		tmp = max(width_in_mb, height_in_mb);
500 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
501 		break;
502 
503 	case 3: /* MPEG2 */
504 
505 		/* reference picture buffer */
506 		min_dpb_size = image_size * 3;
507 		break;
508 
509 	case 4: /* MPEG4 */
510 
511 		/* reference picture buffer */
512 		min_dpb_size = image_size * 3;
513 
514 		/* CM */
515 		min_dpb_size += width_in_mb * height_in_mb * 64;
516 
517 		/* IT surface buffer */
518 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
519 		break;
520 
521 	case 7: /* H264 Perf */
522 		switch(level) {
523 		case 30:
524 			num_dpb_buffer = 8100 / fs_in_mb;
525 			break;
526 		case 31:
527 			num_dpb_buffer = 18000 / fs_in_mb;
528 			break;
529 		case 32:
530 			num_dpb_buffer = 20480 / fs_in_mb;
531 			break;
532 		case 41:
533 			num_dpb_buffer = 32768 / fs_in_mb;
534 			break;
535 		case 42:
536 			num_dpb_buffer = 34816 / fs_in_mb;
537 			break;
538 		case 50:
539 			num_dpb_buffer = 110400 / fs_in_mb;
540 			break;
541 		case 51:
542 			num_dpb_buffer = 184320 / fs_in_mb;
543 			break;
544 		default:
545 			num_dpb_buffer = 184320 / fs_in_mb;
546 			break;
547 		}
548 		num_dpb_buffer++;
549 		if (num_dpb_buffer > 17)
550 			num_dpb_buffer = 17;
551 
552 		/* reference picture buffer */
553 		min_dpb_size = image_size * num_dpb_buffer;
554 
555 		if (!adev->uvd.use_ctx_buf){
556 			/* macroblock context buffer */
557 			min_dpb_size +=
558 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
559 
560 			/* IT surface buffer */
561 			min_dpb_size += width_in_mb * height_in_mb * 32;
562 		} else {
563 			/* macroblock context buffer */
564 			min_ctx_size =
565 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
566 		}
567 		break;
568 
569 	case 16: /* H265 */
570 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
571 		image_size = ALIGN(image_size, 256);
572 
573 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
574 		min_dpb_size = image_size * num_dpb_buffer;
575 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
576 					   * 16 * num_dpb_buffer + 52 * 1024;
577 		break;
578 
579 	default:
580 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
581 		return -EINVAL;
582 	}
583 
584 	if (width > pitch) {
585 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
586 		return -EINVAL;
587 	}
588 
589 	if (dpb_size < min_dpb_size) {
590 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
591 			  dpb_size, min_dpb_size);
592 		return -EINVAL;
593 	}
594 
595 	buf_sizes[0x1] = dpb_size;
596 	buf_sizes[0x2] = image_size;
597 	buf_sizes[0x4] = min_ctx_size;
598 	return 0;
599 }
600 
601 /**
602  * amdgpu_uvd_cs_msg - handle UVD message
603  *
604  * @ctx: UVD parser context
605  * @bo: buffer object containing the message
606  * @offset: offset into the buffer object
607  *
608  * Peek into the UVD message and extract the session id.
609  * Make sure that we don't open up to many sessions.
610  */
611 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
612 			     struct amdgpu_bo *bo, unsigned offset)
613 {
614 	struct amdgpu_device *adev = ctx->parser->adev;
615 	int32_t *msg, msg_type, handle;
616 	void *ptr;
617 	long r;
618 	int i;
619 
620 	if (offset & 0x3F) {
621 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
622 		return -EINVAL;
623 	}
624 
625 	r = amdgpu_bo_kmap(bo, &ptr);
626 	if (r) {
627 		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
628 		return r;
629 	}
630 
631 	msg = ptr + offset;
632 
633 	msg_type = msg[1];
634 	handle = msg[2];
635 
636 	if (handle == 0) {
637 		DRM_ERROR("Invalid UVD handle!\n");
638 		return -EINVAL;
639 	}
640 
641 	switch (msg_type) {
642 	case 0:
643 		/* it's a create msg, calc image size (width * height) */
644 		amdgpu_bo_kunmap(bo);
645 
646 		/* try to alloc a new handle */
647 		for (i = 0; i < adev->uvd.max_handles; ++i) {
648 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
649 				DRM_ERROR("Handle 0x%x already in use!\n", handle);
650 				return -EINVAL;
651 			}
652 
653 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
654 				adev->uvd.filp[i] = ctx->parser->filp;
655 				return 0;
656 			}
657 		}
658 
659 		DRM_ERROR("No more free UVD handles!\n");
660 		return -ENOSPC;
661 
662 	case 1:
663 		/* it's a decode msg, calc buffer sizes */
664 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
665 		amdgpu_bo_kunmap(bo);
666 		if (r)
667 			return r;
668 
669 		/* validate the handle */
670 		for (i = 0; i < adev->uvd.max_handles; ++i) {
671 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
672 				if (adev->uvd.filp[i] != ctx->parser->filp) {
673 					DRM_ERROR("UVD handle collision detected!\n");
674 					return -EINVAL;
675 				}
676 				return 0;
677 			}
678 		}
679 
680 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
681 		return -ENOENT;
682 
683 	case 2:
684 		/* it's a destroy msg, free the handle */
685 		for (i = 0; i < adev->uvd.max_handles; ++i)
686 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
687 		amdgpu_bo_kunmap(bo);
688 		return 0;
689 
690 	default:
691 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
692 		return -EINVAL;
693 	}
694 	BUG();
695 	return -EINVAL;
696 }
697 
698 /**
699  * amdgpu_uvd_cs_pass2 - second parsing round
700  *
701  * @ctx: UVD parser context
702  *
703  * Patch buffer addresses, make sure buffer sizes are correct.
704  */
705 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
706 {
707 	struct amdgpu_bo_va_mapping *mapping;
708 	struct amdgpu_bo *bo;
709 	uint32_t cmd;
710 	uint64_t start, end;
711 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
712 	int r;
713 
714 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
715 	if (mapping == NULL) {
716 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
717 		return -EINVAL;
718 	}
719 
720 	start = amdgpu_bo_gpu_offset(bo);
721 
722 	end = (mapping->it.last + 1 - mapping->it.start);
723 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
724 
725 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
726 	start += addr;
727 
728 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
729 			    lower_32_bits(start));
730 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
731 			    upper_32_bits(start));
732 
733 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
734 	if (cmd < 0x4) {
735 		if ((end - start) < ctx->buf_sizes[cmd]) {
736 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
737 				  (unsigned)(end - start),
738 				  ctx->buf_sizes[cmd]);
739 			return -EINVAL;
740 		}
741 
742 	} else if (cmd == 0x206) {
743 		if ((end - start) < ctx->buf_sizes[4]) {
744 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
745 					  (unsigned)(end - start),
746 					  ctx->buf_sizes[4]);
747 			return -EINVAL;
748 		}
749 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
750 		DRM_ERROR("invalid UVD command %X!\n", cmd);
751 		return -EINVAL;
752 	}
753 
754 	if (!ctx->parser->adev->uvd.address_64_bit) {
755 		if ((start >> 28) != ((end - 1) >> 28)) {
756 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
757 				  start, end);
758 			return -EINVAL;
759 		}
760 
761 		if ((cmd == 0 || cmd == 0x3) &&
762 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
763 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
764 				  start, end);
765 			return -EINVAL;
766 		}
767 	}
768 
769 	if (cmd == 0) {
770 		ctx->has_msg_cmd = true;
771 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
772 		if (r)
773 			return r;
774 	} else if (!ctx->has_msg_cmd) {
775 		DRM_ERROR("Message needed before other commands are send!\n");
776 		return -EINVAL;
777 	}
778 
779 	return 0;
780 }
781 
782 /**
783  * amdgpu_uvd_cs_reg - parse register writes
784  *
785  * @ctx: UVD parser context
786  * @cb: callback function
787  *
788  * Parse the register writes, call cb on each complete command.
789  */
790 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
791 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
792 {
793 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
794 	int i, r;
795 
796 	ctx->idx++;
797 	for (i = 0; i <= ctx->count; ++i) {
798 		unsigned reg = ctx->reg + i;
799 
800 		if (ctx->idx >= ib->length_dw) {
801 			DRM_ERROR("Register command after end of CS!\n");
802 			return -EINVAL;
803 		}
804 
805 		switch (reg) {
806 		case mmUVD_GPCOM_VCPU_DATA0:
807 			ctx->data0 = ctx->idx;
808 			break;
809 		case mmUVD_GPCOM_VCPU_DATA1:
810 			ctx->data1 = ctx->idx;
811 			break;
812 		case mmUVD_GPCOM_VCPU_CMD:
813 			r = cb(ctx);
814 			if (r)
815 				return r;
816 			break;
817 		case mmUVD_ENGINE_CNTL:
818 		case mmUVD_NO_OP:
819 			break;
820 		default:
821 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
822 			return -EINVAL;
823 		}
824 		ctx->idx++;
825 	}
826 	return 0;
827 }
828 
829 /**
830  * amdgpu_uvd_cs_packets - parse UVD packets
831  *
832  * @ctx: UVD parser context
833  * @cb: callback function
834  *
835  * Parse the command stream packets.
836  */
837 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
838 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
839 {
840 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
841 	int r;
842 
843 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
844 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
845 		unsigned type = CP_PACKET_GET_TYPE(cmd);
846 		switch (type) {
847 		case PACKET_TYPE0:
848 			ctx->reg = CP_PACKET0_GET_REG(cmd);
849 			ctx->count = CP_PACKET_GET_COUNT(cmd);
850 			r = amdgpu_uvd_cs_reg(ctx, cb);
851 			if (r)
852 				return r;
853 			break;
854 		case PACKET_TYPE2:
855 			++ctx->idx;
856 			break;
857 		default:
858 			DRM_ERROR("Unknown packet type %d !\n", type);
859 			return -EINVAL;
860 		}
861 	}
862 	return 0;
863 }
864 
865 /**
866  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
867  *
868  * @parser: Command submission parser context
869  *
870  * Parse the command stream, patch in addresses as necessary.
871  */
872 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
873 {
874 	struct amdgpu_uvd_cs_ctx ctx = {};
875 	unsigned buf_sizes[] = {
876 		[0x00000000]	=	2048,
877 		[0x00000001]	=	0xFFFFFFFF,
878 		[0x00000002]	=	0xFFFFFFFF,
879 		[0x00000003]	=	2048,
880 		[0x00000004]	=	0xFFFFFFFF,
881 	};
882 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
883 	int r;
884 
885 	parser->job->vm = NULL;
886 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
887 
888 	if (ib->length_dw % 16) {
889 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
890 			  ib->length_dw);
891 		return -EINVAL;
892 	}
893 
894 	r = amdgpu_cs_sysvm_access_required(parser);
895 	if (r)
896 		return r;
897 
898 	ctx.parser = parser;
899 	ctx.buf_sizes = buf_sizes;
900 	ctx.ib_idx = ib_idx;
901 
902 	/* first round only required on chips without UVD 64 bit address support */
903 	if (!parser->adev->uvd.address_64_bit) {
904 		/* first round, make sure the buffers are actually in the UVD segment */
905 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
906 		if (r)
907 			return r;
908 	}
909 
910 	/* second round, patch buffer addresses into the command stream */
911 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
912 	if (r)
913 		return r;
914 
915 	if (!ctx.has_msg_cmd) {
916 		DRM_ERROR("UVD-IBs need a msg command!\n");
917 		return -EINVAL;
918 	}
919 
920 	return 0;
921 }
922 
923 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
924 			       bool direct, struct dma_fence **fence)
925 {
926 	struct ttm_validate_buffer tv;
927 	struct ww_acquire_ctx ticket;
928 	struct list_head head;
929 	struct amdgpu_job *job;
930 	struct amdgpu_ib *ib;
931 	struct dma_fence *f = NULL;
932 	struct amdgpu_device *adev = ring->adev;
933 	uint64_t addr;
934 	int i, r;
935 
936 	memset(&tv, 0, sizeof(tv));
937 	tv.bo = &bo->tbo;
938 
939 	INIT_LIST_HEAD(&head);
940 	list_add(&tv.head, &head);
941 
942 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
943 	if (r)
944 		return r;
945 
946 	if (!ring->adev->uvd.address_64_bit) {
947 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
948 		amdgpu_uvd_force_into_uvd_segment(bo);
949 	}
950 
951 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
952 	if (r)
953 		goto err;
954 
955 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
956 	if (r)
957 		goto err;
958 
959 	ib = &job->ibs[0];
960 	addr = amdgpu_bo_gpu_offset(bo);
961 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
962 	ib->ptr[1] = addr;
963 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
964 	ib->ptr[3] = addr >> 32;
965 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
966 	ib->ptr[5] = 0;
967 	for (i = 6; i < 16; i += 2) {
968 		ib->ptr[i] = PACKET0(mmUVD_NO_OP, 0);
969 		ib->ptr[i+1] = 0;
970 	}
971 	ib->length_dw = 16;
972 
973 	if (direct) {
974 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, NULL, &f);
975 		job->fence = dma_fence_get(f);
976 		if (r)
977 			goto err_free;
978 
979 		amdgpu_job_free(job);
980 	} else {
981 		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
982 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
983 		if (r)
984 			goto err_free;
985 	}
986 
987 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
988 
989 	if (fence)
990 		*fence = dma_fence_get(f);
991 	amdgpu_bo_unref(&bo);
992 	dma_fence_put(f);
993 
994 	return 0;
995 
996 err_free:
997 	amdgpu_job_free(job);
998 
999 err:
1000 	ttm_eu_backoff_reservation(&ticket, &head);
1001 	return r;
1002 }
1003 
1004 /* multiple fence commands without any stream commands in between can
1005    crash the vcpu so just try to emmit a dummy create/destroy msg to
1006    avoid this */
1007 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1008 			      struct dma_fence **fence)
1009 {
1010 	struct amdgpu_device *adev = ring->adev;
1011 	struct amdgpu_bo *bo;
1012 	uint32_t *msg;
1013 	int r, i;
1014 
1015 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1016 			     AMDGPU_GEM_DOMAIN_VRAM,
1017 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1018 			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1019 			     NULL, NULL, &bo);
1020 	if (r)
1021 		return r;
1022 
1023 	r = amdgpu_bo_reserve(bo, false);
1024 	if (r) {
1025 		amdgpu_bo_unref(&bo);
1026 		return r;
1027 	}
1028 
1029 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1030 	if (r) {
1031 		amdgpu_bo_unreserve(bo);
1032 		amdgpu_bo_unref(&bo);
1033 		return r;
1034 	}
1035 
1036 	/* stitch together an UVD create msg */
1037 	msg[0] = cpu_to_le32(0x00000de4);
1038 	msg[1] = cpu_to_le32(0x00000000);
1039 	msg[2] = cpu_to_le32(handle);
1040 	msg[3] = cpu_to_le32(0x00000000);
1041 	msg[4] = cpu_to_le32(0x00000000);
1042 	msg[5] = cpu_to_le32(0x00000000);
1043 	msg[6] = cpu_to_le32(0x00000000);
1044 	msg[7] = cpu_to_le32(0x00000780);
1045 	msg[8] = cpu_to_le32(0x00000440);
1046 	msg[9] = cpu_to_le32(0x00000000);
1047 	msg[10] = cpu_to_le32(0x01b37000);
1048 	for (i = 11; i < 1024; ++i)
1049 		msg[i] = cpu_to_le32(0x0);
1050 
1051 	amdgpu_bo_kunmap(bo);
1052 	amdgpu_bo_unreserve(bo);
1053 
1054 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1055 }
1056 
1057 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1058 			       bool direct, struct dma_fence **fence)
1059 {
1060 	struct amdgpu_device *adev = ring->adev;
1061 	struct amdgpu_bo *bo;
1062 	uint32_t *msg;
1063 	int r, i;
1064 
1065 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1066 			     AMDGPU_GEM_DOMAIN_VRAM,
1067 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1068 			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1069 			     NULL, NULL, &bo);
1070 	if (r)
1071 		return r;
1072 
1073 	r = amdgpu_bo_reserve(bo, false);
1074 	if (r) {
1075 		amdgpu_bo_unref(&bo);
1076 		return r;
1077 	}
1078 
1079 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1080 	if (r) {
1081 		amdgpu_bo_unreserve(bo);
1082 		amdgpu_bo_unref(&bo);
1083 		return r;
1084 	}
1085 
1086 	/* stitch together an UVD destroy msg */
1087 	msg[0] = cpu_to_le32(0x00000de4);
1088 	msg[1] = cpu_to_le32(0x00000002);
1089 	msg[2] = cpu_to_le32(handle);
1090 	msg[3] = cpu_to_le32(0x00000000);
1091 	for (i = 4; i < 1024; ++i)
1092 		msg[i] = cpu_to_le32(0x0);
1093 
1094 	amdgpu_bo_kunmap(bo);
1095 	amdgpu_bo_unreserve(bo);
1096 
1097 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1098 }
1099 
1100 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1101 {
1102 	struct amdgpu_device *adev =
1103 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1104 	unsigned fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1105 
1106 	if (fences == 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, UVD_IDLE_TIMEOUT);
1114 	}
1115 }
1116 
1117 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1118 {
1119 	struct amdgpu_device *adev = ring->adev;
1120 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1121 
1122 	if (set_clocks) {
1123 		if (adev->pm.dpm_enabled) {
1124 			amdgpu_dpm_enable_uvd(adev, true);
1125 		} else {
1126 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1127 		}
1128 	}
1129 }
1130 
1131 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1132 {
1133 	schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1134 }
1135 
1136 /**
1137  * amdgpu_uvd_ring_test_ib - test ib execution
1138  *
1139  * @ring: amdgpu_ring pointer
1140  *
1141  * Test if we can successfully execute an IB
1142  */
1143 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1144 {
1145 	struct dma_fence *fence;
1146 	long r;
1147 
1148 	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1149 	if (r) {
1150 		DRM_ERROR("amdgpu: failed to get create msg (%ld).\n", r);
1151 		goto error;
1152 	}
1153 
1154 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1155 	if (r) {
1156 		DRM_ERROR("amdgpu: failed to get destroy ib (%ld).\n", r);
1157 		goto error;
1158 	}
1159 
1160 	r = dma_fence_wait_timeout(fence, false, timeout);
1161 	if (r == 0) {
1162 		DRM_ERROR("amdgpu: IB test timed out.\n");
1163 		r = -ETIMEDOUT;
1164 	} else if (r < 0) {
1165 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1166 	} else {
1167 		DRM_INFO("ib test on ring %d succeeded\n",  ring->idx);
1168 		r = 0;
1169 	}
1170 
1171 	dma_fence_put(fence);
1172 
1173 error:
1174 	return r;
1175 }
1176