xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
69 
70 /**
71  * amdgpu_uvd_cs_ctx - Command submission parser context
72  *
73  * Used for emulating virtual memory support on UVD 4.2.
74  */
75 struct amdgpu_uvd_cs_ctx {
76 	struct amdgpu_cs_parser *parser;
77 	unsigned reg, count;
78 	unsigned data0, data1;
79 	unsigned idx;
80 	unsigned ib_idx;
81 
82 	/* does the IB has a msg command */
83 	bool has_msg_cmd;
84 
85 	/* minimum buffer sizes */
86 	unsigned *buf_sizes;
87 };
88 
89 #ifdef CONFIG_DRM_AMDGPU_CIK
90 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
91 MODULE_FIRMWARE(FIRMWARE_KABINI);
92 MODULE_FIRMWARE(FIRMWARE_KAVERI);
93 MODULE_FIRMWARE(FIRMWARE_HAWAII);
94 MODULE_FIRMWARE(FIRMWARE_MULLINS);
95 #endif
96 MODULE_FIRMWARE(FIRMWARE_TONGA);
97 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
98 MODULE_FIRMWARE(FIRMWARE_FIJI);
99 MODULE_FIRMWARE(FIRMWARE_STONEY);
100 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
101 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
102 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
103 
104 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
105 
106 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
107 {
108 	struct amdgpu_ring *ring;
109 	struct amd_sched_rq *rq;
110 	unsigned long bo_size;
111 	const char *fw_name;
112 	const struct common_firmware_header *hdr;
113 	unsigned version_major, version_minor, family_id;
114 	int i, r;
115 
116 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
117 
118 	switch (adev->asic_type) {
119 #ifdef CONFIG_DRM_AMDGPU_CIK
120 	case CHIP_BONAIRE:
121 		fw_name = FIRMWARE_BONAIRE;
122 		break;
123 	case CHIP_KABINI:
124 		fw_name = FIRMWARE_KABINI;
125 		break;
126 	case CHIP_KAVERI:
127 		fw_name = FIRMWARE_KAVERI;
128 		break;
129 	case CHIP_HAWAII:
130 		fw_name = FIRMWARE_HAWAII;
131 		break;
132 	case CHIP_MULLINS:
133 		fw_name = FIRMWARE_MULLINS;
134 		break;
135 #endif
136 	case CHIP_TONGA:
137 		fw_name = FIRMWARE_TONGA;
138 		break;
139 	case CHIP_FIJI:
140 		fw_name = FIRMWARE_FIJI;
141 		break;
142 	case CHIP_CARRIZO:
143 		fw_name = FIRMWARE_CARRIZO;
144 		break;
145 	case CHIP_STONEY:
146 		fw_name = FIRMWARE_STONEY;
147 		break;
148 	case CHIP_POLARIS10:
149 		fw_name = FIRMWARE_POLARIS10;
150 		break;
151 	case CHIP_POLARIS11:
152 		fw_name = FIRMWARE_POLARIS11;
153 		break;
154 	case CHIP_POLARIS12:
155 		fw_name = FIRMWARE_POLARIS12;
156 		break;
157 	default:
158 		return -EINVAL;
159 	}
160 
161 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
162 	if (r) {
163 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
164 			fw_name);
165 		return r;
166 	}
167 
168 	r = amdgpu_ucode_validate(adev->uvd.fw);
169 	if (r) {
170 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
171 			fw_name);
172 		release_firmware(adev->uvd.fw);
173 		adev->uvd.fw = NULL;
174 		return r;
175 	}
176 
177 	/* Set the default UVD handles that the firmware can handle */
178 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
179 
180 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
181 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
182 	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
183 	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
184 	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
185 		version_major, version_minor, family_id);
186 
187 	/*
188 	 * Limit the number of UVD handles depending on microcode major
189 	 * and minor versions. The firmware version which has 40 UVD
190 	 * instances support is 1.80. So all subsequent versions should
191 	 * also have the same support.
192 	 */
193 	if ((version_major > 0x01) ||
194 	    ((version_major == 0x01) && (version_minor >= 0x50)))
195 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
196 
197 	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
198 				(family_id << 8));
199 
200 	if ((adev->asic_type == CHIP_POLARIS10 ||
201 	     adev->asic_type == CHIP_POLARIS11) &&
202 	    (adev->uvd.fw_version < FW_1_66_16))
203 		DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
204 			  version_major, version_minor);
205 
206 	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
207 		  +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
208 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
209 	r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
210 				    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.vcpu_bo,
211 				    &adev->uvd.gpu_addr, &adev->uvd.cpu_addr);
212 	if (r) {
213 		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
214 		return r;
215 	}
216 
217 	ring = &adev->uvd.ring;
218 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
219 	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
220 				  rq, amdgpu_sched_jobs);
221 	if (r != 0) {
222 		DRM_ERROR("Failed setting up UVD run queue.\n");
223 		return r;
224 	}
225 
226 	for (i = 0; i < adev->uvd.max_handles; ++i) {
227 		atomic_set(&adev->uvd.handles[i], 0);
228 		adev->uvd.filp[i] = NULL;
229 	}
230 
231 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
232 	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
233 		adev->uvd.address_64_bit = true;
234 
235 	switch (adev->asic_type) {
236 	case CHIP_TONGA:
237 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
238 		break;
239 	case CHIP_CARRIZO:
240 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
241 		break;
242 	case CHIP_FIJI:
243 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
244 		break;
245 	case CHIP_STONEY:
246 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
247 		break;
248 	default:
249 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
250 	}
251 
252 	return 0;
253 }
254 
255 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
256 {
257 	kfree(adev->uvd.saved_bo);
258 
259 	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
260 
261 	amdgpu_bo_free_kernel(&adev->uvd.vcpu_bo,
262 			      &adev->uvd.gpu_addr,
263 			      (void **)&adev->uvd.cpu_addr);
264 
265 	amdgpu_ring_fini(&adev->uvd.ring);
266 
267 	release_firmware(adev->uvd.fw);
268 
269 	return 0;
270 }
271 
272 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
273 {
274 	unsigned size;
275 	void *ptr;
276 	int i;
277 
278 	if (adev->uvd.vcpu_bo == NULL)
279 		return 0;
280 
281 	for (i = 0; i < adev->uvd.max_handles; ++i)
282 		if (atomic_read(&adev->uvd.handles[i]))
283 			break;
284 
285 	if (i == AMDGPU_MAX_UVD_HANDLES)
286 		return 0;
287 
288 	cancel_delayed_work_sync(&adev->uvd.idle_work);
289 
290 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
291 	ptr = adev->uvd.cpu_addr;
292 
293 	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
294 	if (!adev->uvd.saved_bo)
295 		return -ENOMEM;
296 
297 	memcpy_fromio(adev->uvd.saved_bo, ptr, size);
298 
299 	return 0;
300 }
301 
302 int amdgpu_uvd_resume(struct amdgpu_device *adev)
303 {
304 	unsigned size;
305 	void *ptr;
306 
307 	if (adev->uvd.vcpu_bo == NULL)
308 		return -EINVAL;
309 
310 	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
311 	ptr = adev->uvd.cpu_addr;
312 
313 	if (adev->uvd.saved_bo != NULL) {
314 		memcpy_toio(ptr, adev->uvd.saved_bo, size);
315 		kfree(adev->uvd.saved_bo);
316 		adev->uvd.saved_bo = NULL;
317 	} else {
318 		const struct common_firmware_header *hdr;
319 		unsigned offset;
320 
321 		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
322 		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
323 		memcpy_toio(adev->uvd.cpu_addr, adev->uvd.fw->data + offset,
324 			    le32_to_cpu(hdr->ucode_size_bytes));
325 		size -= le32_to_cpu(hdr->ucode_size_bytes);
326 		ptr += le32_to_cpu(hdr->ucode_size_bytes);
327 		memset_io(ptr, 0, size);
328 	}
329 
330 	return 0;
331 }
332 
333 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
334 {
335 	struct amdgpu_ring *ring = &adev->uvd.ring;
336 	int i, r;
337 
338 	for (i = 0; i < adev->uvd.max_handles; ++i) {
339 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
340 		if (handle != 0 && adev->uvd.filp[i] == filp) {
341 			struct dma_fence *fence;
342 
343 			r = amdgpu_uvd_get_destroy_msg(ring, handle,
344 						       false, &fence);
345 			if (r) {
346 				DRM_ERROR("Error destroying UVD (%d)!\n", r);
347 				continue;
348 			}
349 
350 			dma_fence_wait(fence, false);
351 			dma_fence_put(fence);
352 
353 			adev->uvd.filp[i] = NULL;
354 			atomic_set(&adev->uvd.handles[i], 0);
355 		}
356 	}
357 }
358 
359 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
360 {
361 	int i;
362 	for (i = 0; i < abo->placement.num_placement; ++i) {
363 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
364 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
365 	}
366 }
367 
368 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
369 {
370 	uint32_t lo, hi;
371 	uint64_t addr;
372 
373 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
374 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
375 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
376 
377 	return addr;
378 }
379 
380 /**
381  * amdgpu_uvd_cs_pass1 - first parsing round
382  *
383  * @ctx: UVD parser context
384  *
385  * Make sure UVD message and feedback buffers are in VRAM and
386  * nobody is violating an 256MB boundary.
387  */
388 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
389 {
390 	struct amdgpu_bo_va_mapping *mapping;
391 	struct amdgpu_bo *bo;
392 	uint32_t cmd;
393 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
394 	int r = 0;
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->uvd.use_ctx_buf){
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 -ENOSPC;
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;
715 	uint64_t start, end;
716 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
717 	int r;
718 
719 	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
720 	if (mapping == NULL) {
721 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
722 		return -EINVAL;
723 	}
724 
725 	start = amdgpu_bo_gpu_offset(bo);
726 
727 	end = (mapping->it.last + 1 - mapping->it.start);
728 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
729 
730 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
731 	start += addr;
732 
733 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
734 			    lower_32_bits(start));
735 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
736 			    upper_32_bits(start));
737 
738 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
739 	if (cmd < 0x4) {
740 		if ((end - start) < ctx->buf_sizes[cmd]) {
741 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
742 				  (unsigned)(end - start),
743 				  ctx->buf_sizes[cmd]);
744 			return -EINVAL;
745 		}
746 
747 	} else if (cmd == 0x206) {
748 		if ((end - start) < ctx->buf_sizes[4]) {
749 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
750 					  (unsigned)(end - start),
751 					  ctx->buf_sizes[4]);
752 			return -EINVAL;
753 		}
754 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
755 		DRM_ERROR("invalid UVD command %X!\n", cmd);
756 		return -EINVAL;
757 	}
758 
759 	if (!ctx->parser->adev->uvd.address_64_bit) {
760 		if ((start >> 28) != ((end - 1) >> 28)) {
761 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
762 				  start, end);
763 			return -EINVAL;
764 		}
765 
766 		if ((cmd == 0 || cmd == 0x3) &&
767 		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
768 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
769 				  start, end);
770 			return -EINVAL;
771 		}
772 	}
773 
774 	if (cmd == 0) {
775 		ctx->has_msg_cmd = true;
776 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
777 		if (r)
778 			return r;
779 	} else if (!ctx->has_msg_cmd) {
780 		DRM_ERROR("Message needed before other commands are send!\n");
781 		return -EINVAL;
782 	}
783 
784 	return 0;
785 }
786 
787 /**
788  * amdgpu_uvd_cs_reg - parse register writes
789  *
790  * @ctx: UVD parser context
791  * @cb: callback function
792  *
793  * Parse the register writes, call cb on each complete command.
794  */
795 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
796 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
797 {
798 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
799 	int i, r;
800 
801 	ctx->idx++;
802 	for (i = 0; i <= ctx->count; ++i) {
803 		unsigned reg = ctx->reg + i;
804 
805 		if (ctx->idx >= ib->length_dw) {
806 			DRM_ERROR("Register command after end of CS!\n");
807 			return -EINVAL;
808 		}
809 
810 		switch (reg) {
811 		case mmUVD_GPCOM_VCPU_DATA0:
812 			ctx->data0 = ctx->idx;
813 			break;
814 		case mmUVD_GPCOM_VCPU_DATA1:
815 			ctx->data1 = ctx->idx;
816 			break;
817 		case mmUVD_GPCOM_VCPU_CMD:
818 			r = cb(ctx);
819 			if (r)
820 				return r;
821 			break;
822 		case mmUVD_ENGINE_CNTL:
823 		case mmUVD_NO_OP:
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 	parser->job->vm = NULL;
891 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
892 
893 	if (ib->length_dw % 16) {
894 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
895 			  ib->length_dw);
896 		return -EINVAL;
897 	}
898 
899 	r = amdgpu_cs_sysvm_access_required(parser);
900 	if (r)
901 		return r;
902 
903 	ctx.parser = parser;
904 	ctx.buf_sizes = buf_sizes;
905 	ctx.ib_idx = ib_idx;
906 
907 	/* first round only required on chips without UVD 64 bit address support */
908 	if (!parser->adev->uvd.address_64_bit) {
909 		/* first round, make sure the buffers are actually in the UVD segment */
910 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
911 		if (r)
912 			return r;
913 	}
914 
915 	/* second round, patch buffer addresses into the command stream */
916 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
917 	if (r)
918 		return r;
919 
920 	if (!ctx.has_msg_cmd) {
921 		DRM_ERROR("UVD-IBs need a msg command!\n");
922 		return -EINVAL;
923 	}
924 
925 	return 0;
926 }
927 
928 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
929 			       bool direct, struct dma_fence **fence)
930 {
931 	struct ttm_validate_buffer tv;
932 	struct ww_acquire_ctx ticket;
933 	struct list_head head;
934 	struct amdgpu_job *job;
935 	struct amdgpu_ib *ib;
936 	struct dma_fence *f = NULL;
937 	struct amdgpu_device *adev = ring->adev;
938 	uint64_t addr;
939 	int i, r;
940 
941 	memset(&tv, 0, sizeof(tv));
942 	tv.bo = &bo->tbo;
943 
944 	INIT_LIST_HEAD(&head);
945 	list_add(&tv.head, &head);
946 
947 	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
948 	if (r)
949 		return r;
950 
951 	if (!ring->adev->uvd.address_64_bit) {
952 		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
953 		amdgpu_uvd_force_into_uvd_segment(bo);
954 	}
955 
956 	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
957 	if (r)
958 		goto err;
959 
960 	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
961 	if (r)
962 		goto err;
963 
964 	ib = &job->ibs[0];
965 	addr = amdgpu_bo_gpu_offset(bo);
966 	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
967 	ib->ptr[1] = addr;
968 	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
969 	ib->ptr[3] = addr >> 32;
970 	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
971 	ib->ptr[5] = 0;
972 	for (i = 6; i < 16; i += 2) {
973 		ib->ptr[i] = PACKET0(mmUVD_NO_OP, 0);
974 		ib->ptr[i+1] = 0;
975 	}
976 	ib->length_dw = 16;
977 
978 	if (direct) {
979 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
980 		job->fence = dma_fence_get(f);
981 		if (r)
982 			goto err_free;
983 
984 		amdgpu_job_free(job);
985 	} else {
986 		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
987 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
988 		if (r)
989 			goto err_free;
990 	}
991 
992 	ttm_eu_fence_buffer_objects(&ticket, &head, f);
993 
994 	if (fence)
995 		*fence = dma_fence_get(f);
996 	amdgpu_bo_unref(&bo);
997 	dma_fence_put(f);
998 
999 	return 0;
1000 
1001 err_free:
1002 	amdgpu_job_free(job);
1003 
1004 err:
1005 	ttm_eu_backoff_reservation(&ticket, &head);
1006 	return r;
1007 }
1008 
1009 /* multiple fence commands without any stream commands in between can
1010    crash the vcpu so just try to emmit a dummy create/destroy msg to
1011    avoid this */
1012 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1013 			      struct dma_fence **fence)
1014 {
1015 	struct amdgpu_device *adev = ring->adev;
1016 	struct amdgpu_bo *bo;
1017 	uint32_t *msg;
1018 	int r, i;
1019 
1020 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1021 			     AMDGPU_GEM_DOMAIN_VRAM,
1022 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1023 			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1024 			     NULL, NULL, &bo);
1025 	if (r)
1026 		return r;
1027 
1028 	r = amdgpu_bo_reserve(bo, false);
1029 	if (r) {
1030 		amdgpu_bo_unref(&bo);
1031 		return r;
1032 	}
1033 
1034 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1035 	if (r) {
1036 		amdgpu_bo_unreserve(bo);
1037 		amdgpu_bo_unref(&bo);
1038 		return r;
1039 	}
1040 
1041 	/* stitch together an UVD create msg */
1042 	msg[0] = cpu_to_le32(0x00000de4);
1043 	msg[1] = cpu_to_le32(0x00000000);
1044 	msg[2] = cpu_to_le32(handle);
1045 	msg[3] = cpu_to_le32(0x00000000);
1046 	msg[4] = cpu_to_le32(0x00000000);
1047 	msg[5] = cpu_to_le32(0x00000000);
1048 	msg[6] = cpu_to_le32(0x00000000);
1049 	msg[7] = cpu_to_le32(0x00000780);
1050 	msg[8] = cpu_to_le32(0x00000440);
1051 	msg[9] = cpu_to_le32(0x00000000);
1052 	msg[10] = cpu_to_le32(0x01b37000);
1053 	for (i = 11; i < 1024; ++i)
1054 		msg[i] = cpu_to_le32(0x0);
1055 
1056 	amdgpu_bo_kunmap(bo);
1057 	amdgpu_bo_unreserve(bo);
1058 
1059 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1060 }
1061 
1062 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1063 			       bool direct, struct dma_fence **fence)
1064 {
1065 	struct amdgpu_device *adev = ring->adev;
1066 	struct amdgpu_bo *bo;
1067 	uint32_t *msg;
1068 	int r, i;
1069 
1070 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1071 			     AMDGPU_GEM_DOMAIN_VRAM,
1072 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1073 			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1074 			     NULL, NULL, &bo);
1075 	if (r)
1076 		return r;
1077 
1078 	r = amdgpu_bo_reserve(bo, false);
1079 	if (r) {
1080 		amdgpu_bo_unref(&bo);
1081 		return r;
1082 	}
1083 
1084 	r = amdgpu_bo_kmap(bo, (void **)&msg);
1085 	if (r) {
1086 		amdgpu_bo_unreserve(bo);
1087 		amdgpu_bo_unref(&bo);
1088 		return r;
1089 	}
1090 
1091 	/* stitch together an UVD destroy msg */
1092 	msg[0] = cpu_to_le32(0x00000de4);
1093 	msg[1] = cpu_to_le32(0x00000002);
1094 	msg[2] = cpu_to_le32(handle);
1095 	msg[3] = cpu_to_le32(0x00000000);
1096 	for (i = 4; i < 1024; ++i)
1097 		msg[i] = cpu_to_le32(0x0);
1098 
1099 	amdgpu_bo_kunmap(bo);
1100 	amdgpu_bo_unreserve(bo);
1101 
1102 	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1103 }
1104 
1105 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1106 {
1107 	struct amdgpu_device *adev =
1108 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1109 	unsigned fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1110 
1111 	if (fences == 0) {
1112 		if (adev->pm.dpm_enabled) {
1113 			amdgpu_dpm_enable_uvd(adev, false);
1114 		} else {
1115 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1116 			/* shutdown the UVD block */
1117 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1118 							    AMD_PG_STATE_GATE);
1119 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1120 							    AMD_CG_STATE_GATE);
1121 		}
1122 	} else {
1123 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1124 	}
1125 }
1126 
1127 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1128 {
1129 	struct amdgpu_device *adev = ring->adev;
1130 	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1131 
1132 	if (set_clocks) {
1133 		if (adev->pm.dpm_enabled) {
1134 			amdgpu_dpm_enable_uvd(adev, true);
1135 		} else {
1136 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1137 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1138 							    AMD_CG_STATE_UNGATE);
1139 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1140 							    AMD_PG_STATE_UNGATE);
1141 		}
1142 	}
1143 }
1144 
1145 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1146 {
1147 	schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1148 }
1149 
1150 /**
1151  * amdgpu_uvd_ring_test_ib - test ib execution
1152  *
1153  * @ring: amdgpu_ring pointer
1154  *
1155  * Test if we can successfully execute an IB
1156  */
1157 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1158 {
1159 	struct dma_fence *fence;
1160 	long r;
1161 
1162 	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1163 	if (r) {
1164 		DRM_ERROR("amdgpu: failed to get create msg (%ld).\n", r);
1165 		goto error;
1166 	}
1167 
1168 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1169 	if (r) {
1170 		DRM_ERROR("amdgpu: failed to get destroy ib (%ld).\n", r);
1171 		goto error;
1172 	}
1173 
1174 	r = dma_fence_wait_timeout(fence, false, timeout);
1175 	if (r == 0) {
1176 		DRM_ERROR("amdgpu: IB test timed out.\n");
1177 		r = -ETIMEDOUT;
1178 	} else if (r < 0) {
1179 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1180 	} else {
1181 		DRM_INFO("ib test on ring %d succeeded\n",  ring->idx);
1182 		r = 0;
1183 	}
1184 
1185 	dma_fence_put(fence);
1186 
1187 error:
1188 	return r;
1189 }
1190 
1191 /**
1192  * amdgpu_uvd_used_handles - returns used UVD handles
1193  *
1194  * @adev: amdgpu_device pointer
1195  *
1196  * Returns the number of UVD handles in use
1197  */
1198 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1199 {
1200 	unsigned i;
1201 	uint32_t used_handles = 0;
1202 
1203 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1204 		/*
1205 		 * Handles can be freed in any order, and not
1206 		 * necessarily linear. So we need to count
1207 		 * all non-zero handles.
1208 		 */
1209 		if (atomic_read(&adev->uvd.handles[i]))
1210 			used_handles++;
1211 	}
1212 
1213 	return used_handles;
1214 }
1215