xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision 4dd10fbbeb9c33709edc89916e19c03c780a0751)
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 
34 #include <drm/drm.h>
35 #include <drm/drm_drv.h>
36 
37 #include "amdgpu.h"
38 #include "amdgpu_pm.h"
39 #include "amdgpu_uvd.h"
40 #include "amdgpu_cs.h"
41 #include "cikd.h"
42 #include "uvd/uvd_4_2_d.h"
43 
44 #include "amdgpu_ras.h"
45 
46 /* 1 second timeout */
47 #define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
48 
49 /* Firmware versions for VI */
50 #define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
51 #define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
52 #define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
53 #define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
54 
55 /* Polaris10/11 firmware version */
56 #define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
57 
58 /* Firmware Names */
59 #ifdef CONFIG_DRM_AMDGPU_SI
60 #define FIRMWARE_TAHITI		"amdgpu/tahiti_uvd.bin"
61 #define FIRMWARE_VERDE		"amdgpu/verde_uvd.bin"
62 #define FIRMWARE_PITCAIRN	"amdgpu/pitcairn_uvd.bin"
63 #define FIRMWARE_OLAND		"amdgpu/oland_uvd.bin"
64 #endif
65 #ifdef CONFIG_DRM_AMDGPU_CIK
66 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_uvd.bin"
67 #define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
68 #define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
69 #define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
70 #define FIRMWARE_MULLINS	"amdgpu/mullins_uvd.bin"
71 #endif
72 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
73 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
74 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
75 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
76 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
77 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
78 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
79 #define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
80 
81 #define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
82 #define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
83 #define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
84 
85 /* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
86 #define UVD_GPCOM_VCPU_CMD		0x03c3
87 #define UVD_GPCOM_VCPU_DATA0	0x03c4
88 #define UVD_GPCOM_VCPU_DATA1	0x03c5
89 #define UVD_NO_OP				0x03ff
90 #define UVD_BASE_SI				0x3800
91 
92 /*
93  * amdgpu_uvd_cs_ctx - Command submission parser context
94  *
95  * Used for emulating virtual memory support on UVD 4.2.
96  */
97 struct amdgpu_uvd_cs_ctx {
98 	struct amdgpu_cs_parser *parser;
99 	unsigned int reg, count;
100 	unsigned int data0, data1;
101 	unsigned int idx;
102 	struct amdgpu_ib *ib;
103 
104 	/* does the IB has a msg command */
105 	bool has_msg_cmd;
106 
107 	/* minimum buffer sizes */
108 	unsigned int *buf_sizes;
109 };
110 
111 #ifdef CONFIG_DRM_AMDGPU_SI
112 MODULE_FIRMWARE(FIRMWARE_TAHITI);
113 MODULE_FIRMWARE(FIRMWARE_VERDE);
114 MODULE_FIRMWARE(FIRMWARE_PITCAIRN);
115 MODULE_FIRMWARE(FIRMWARE_OLAND);
116 #endif
117 #ifdef CONFIG_DRM_AMDGPU_CIK
118 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
119 MODULE_FIRMWARE(FIRMWARE_KABINI);
120 MODULE_FIRMWARE(FIRMWARE_KAVERI);
121 MODULE_FIRMWARE(FIRMWARE_HAWAII);
122 MODULE_FIRMWARE(FIRMWARE_MULLINS);
123 #endif
124 MODULE_FIRMWARE(FIRMWARE_TONGA);
125 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
126 MODULE_FIRMWARE(FIRMWARE_FIJI);
127 MODULE_FIRMWARE(FIRMWARE_STONEY);
128 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
129 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
130 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
131 MODULE_FIRMWARE(FIRMWARE_VEGAM);
132 
133 MODULE_FIRMWARE(FIRMWARE_VEGA10);
134 MODULE_FIRMWARE(FIRMWARE_VEGA12);
135 MODULE_FIRMWARE(FIRMWARE_VEGA20);
136 
137 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
138 static void amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo *abo);
139 
140 static int amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device *adev,
141 					   uint32_t size,
142 					   struct amdgpu_bo **bo_ptr)
143 {
144 	struct ttm_operation_ctx ctx = { true, false };
145 	struct amdgpu_bo *bo = NULL;
146 	void *addr;
147 	int r;
148 
149 	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
150 				      AMDGPU_GEM_DOMAIN_GTT,
151 				      &bo, NULL, &addr);
152 	if (r)
153 		return r;
154 
155 	if (adev->uvd.address_64_bit)
156 		goto succ;
157 
158 	amdgpu_bo_kunmap(bo);
159 	amdgpu_bo_unpin(bo);
160 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
161 	amdgpu_uvd_force_into_vcpu_segment(bo);
162 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
163 	if (r)
164 		goto err;
165 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
166 	if (r)
167 		goto err_pin;
168 	r = amdgpu_bo_kmap(bo, &addr);
169 	if (r)
170 		goto err_kmap;
171 succ:
172 	amdgpu_bo_unreserve(bo);
173 	*bo_ptr = bo;
174 	return 0;
175 err_kmap:
176 	amdgpu_bo_unpin(bo);
177 err_pin:
178 err:
179 	amdgpu_bo_unreserve(bo);
180 	amdgpu_bo_unref(&bo);
181 	return r;
182 }
183 
184 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
185 {
186 	unsigned long bo_size;
187 	const char *fw_name;
188 	const struct common_firmware_header *hdr;
189 	unsigned int family_id;
190 	int i, j, r;
191 	u32 vcpu_bo_domain;
192 
193 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
194 
195 	switch (adev->asic_type) {
196 #ifdef CONFIG_DRM_AMDGPU_SI
197 	case CHIP_TAHITI:
198 		fw_name = FIRMWARE_TAHITI;
199 		break;
200 	case CHIP_VERDE:
201 		fw_name = FIRMWARE_VERDE;
202 		break;
203 	case CHIP_PITCAIRN:
204 		fw_name = FIRMWARE_PITCAIRN;
205 		break;
206 	case CHIP_OLAND:
207 		fw_name = FIRMWARE_OLAND;
208 		break;
209 #endif
210 #ifdef CONFIG_DRM_AMDGPU_CIK
211 	case CHIP_BONAIRE:
212 		fw_name = FIRMWARE_BONAIRE;
213 		break;
214 	case CHIP_KABINI:
215 		fw_name = FIRMWARE_KABINI;
216 		break;
217 	case CHIP_KAVERI:
218 		fw_name = FIRMWARE_KAVERI;
219 		break;
220 	case CHIP_HAWAII:
221 		fw_name = FIRMWARE_HAWAII;
222 		break;
223 	case CHIP_MULLINS:
224 		fw_name = FIRMWARE_MULLINS;
225 		break;
226 #endif
227 	case CHIP_TONGA:
228 		fw_name = FIRMWARE_TONGA;
229 		break;
230 	case CHIP_FIJI:
231 		fw_name = FIRMWARE_FIJI;
232 		break;
233 	case CHIP_CARRIZO:
234 		fw_name = FIRMWARE_CARRIZO;
235 		break;
236 	case CHIP_STONEY:
237 		fw_name = FIRMWARE_STONEY;
238 		break;
239 	case CHIP_POLARIS10:
240 		fw_name = FIRMWARE_POLARIS10;
241 		break;
242 	case CHIP_POLARIS11:
243 		fw_name = FIRMWARE_POLARIS11;
244 		break;
245 	case CHIP_POLARIS12:
246 		fw_name = FIRMWARE_POLARIS12;
247 		break;
248 	case CHIP_VEGA10:
249 		fw_name = FIRMWARE_VEGA10;
250 		break;
251 	case CHIP_VEGA12:
252 		fw_name = FIRMWARE_VEGA12;
253 		break;
254 	case CHIP_VEGAM:
255 		fw_name = FIRMWARE_VEGAM;
256 		break;
257 	case CHIP_VEGA20:
258 		fw_name = FIRMWARE_VEGA20;
259 		break;
260 	default:
261 		return -EINVAL;
262 	}
263 
264 	r = amdgpu_ucode_request(adev, &adev->uvd.fw, AMDGPU_UCODE_REQUIRED, "%s", fw_name);
265 	if (r) {
266 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
267 			fw_name);
268 		amdgpu_ucode_release(&adev->uvd.fw);
269 		return r;
270 	}
271 
272 	/* Set the default UVD handles that the firmware can handle */
273 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
274 
275 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
276 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
277 
278 	if (adev->asic_type < CHIP_VEGA20) {
279 		unsigned int version_major, version_minor;
280 
281 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
282 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
283 		drm_info(adev_to_drm(adev), "Found UVD firmware Version: %u.%u Family ID: %u\n",
284 			version_major, version_minor, family_id);
285 
286 		/*
287 		 * Limit the number of UVD handles depending on microcode major
288 		 * and minor versions. The firmware version which has 40 UVD
289 		 * instances support is 1.80. So all subsequent versions should
290 		 * also have the same support.
291 		 */
292 		if ((version_major > 0x01) ||
293 		    ((version_major == 0x01) && (version_minor >= 0x50)))
294 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
295 
296 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
297 					(family_id << 8));
298 
299 		if ((adev->asic_type == CHIP_POLARIS10 ||
300 		     adev->asic_type == CHIP_POLARIS11) &&
301 		    (adev->uvd.fw_version < FW_1_66_16))
302 			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
303 				  version_major, version_minor);
304 	} else {
305 		unsigned int enc_major, enc_minor, dec_minor;
306 
307 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
308 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
309 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
310 		drm_info(adev_to_drm(adev), "Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
311 			enc_major, enc_minor, dec_minor, family_id);
312 
313 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
314 
315 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
316 	}
317 
318 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
319 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
320 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
321 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
322 
323 	/* UVD 5.0 and newer HW can use 64 bit addressing. */
324 	adev->uvd.address_64_bit =
325 		!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0);
326 
327 	vcpu_bo_domain = AMDGPU_GEM_DOMAIN_VRAM;
328 	if (adev->uvd.address_64_bit)
329 		vcpu_bo_domain |= AMDGPU_GEM_DOMAIN_GTT;
330 
331 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
332 		if (adev->uvd.harvest_config & (1 << j))
333 			continue;
334 
335 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
336 					    vcpu_bo_domain,
337 					    &adev->uvd.inst[j].vcpu_bo,
338 					    &adev->uvd.inst[j].gpu_addr,
339 					    &adev->uvd.inst[j].cpu_addr);
340 		if (r) {
341 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
342 			return r;
343 		}
344 	}
345 
346 	for (i = 0; i < adev->uvd.max_handles; ++i) {
347 		atomic_set(&adev->uvd.handles[i], 0);
348 		adev->uvd.filp[i] = NULL;
349 	}
350 
351 	r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
352 	if (r)
353 		return r;
354 
355 	switch (adev->asic_type) {
356 	case CHIP_TONGA:
357 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
358 		break;
359 	case CHIP_CARRIZO:
360 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
361 		break;
362 	case CHIP_FIJI:
363 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
364 		break;
365 	case CHIP_STONEY:
366 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
367 		break;
368 	default:
369 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
370 	}
371 
372 	return 0;
373 }
374 
375 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
376 {
377 	void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
378 	int i, j;
379 
380 	if (adev->uvd.entity.rq)
381 		drm_sched_entity_destroy(&adev->uvd.entity);
382 
383 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
384 		if (adev->uvd.harvest_config & (1 << j))
385 			continue;
386 		kvfree(adev->uvd.inst[j].saved_bo);
387 
388 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
389 				      &adev->uvd.inst[j].gpu_addr,
390 				      (void **)&adev->uvd.inst[j].cpu_addr);
391 
392 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
393 
394 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
395 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
396 	}
397 	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
398 	amdgpu_ucode_release(&adev->uvd.fw);
399 
400 	return 0;
401 }
402 
403 /**
404  * amdgpu_uvd_entity_init - init entity
405  *
406  * @adev: amdgpu_device pointer
407  * @ring: amdgpu_ring pointer to check
408  *
409  * Initialize the entity used for handle management in the kernel driver.
410  */
411 int amdgpu_uvd_entity_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
412 {
413 	if (ring == &adev->uvd.inst[0].ring) {
414 		struct drm_gpu_scheduler *sched = &ring->sched;
415 		int r;
416 
417 		r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
418 					  &sched, 1, NULL);
419 		if (r) {
420 			DRM_ERROR("Failed setting up UVD kernel entity.\n");
421 			return r;
422 		}
423 	}
424 
425 	return 0;
426 }
427 
428 int amdgpu_uvd_prepare_suspend(struct amdgpu_device *adev)
429 {
430 	unsigned int size;
431 	void *ptr;
432 	int i, j, idx;
433 
434 	cancel_delayed_work_sync(&adev->uvd.idle_work);
435 
436 	/* only valid for physical mode */
437 	if (adev->asic_type < CHIP_POLARIS10) {
438 		for (i = 0; i < adev->uvd.max_handles; ++i)
439 			if (atomic_read(&adev->uvd.handles[i]))
440 				break;
441 
442 		if (i == adev->uvd.max_handles)
443 			return 0;
444 	}
445 
446 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
447 		if (adev->uvd.harvest_config & (1 << j))
448 			continue;
449 		if (adev->uvd.inst[j].vcpu_bo == NULL)
450 			continue;
451 
452 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
453 		ptr = adev->uvd.inst[j].cpu_addr;
454 
455 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
456 		if (!adev->uvd.inst[j].saved_bo)
457 			return -ENOMEM;
458 
459 		if (drm_dev_enter(adev_to_drm(adev), &idx)) {
460 			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
461 			if (amdgpu_ras_intr_triggered())
462 				memset(adev->uvd.inst[j].saved_bo, 0, size);
463 			else
464 				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
465 
466 			drm_dev_exit(idx);
467 		}
468 	}
469 
470 	return 0;
471 }
472 
473 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
474 {
475 	if (amdgpu_ras_intr_triggered())
476 		drm_warn(adev_to_drm(adev),
477 			"UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
478 
479 	return 0;
480 }
481 
482 int amdgpu_uvd_resume(struct amdgpu_device *adev)
483 {
484 	unsigned int size;
485 	void *ptr;
486 	int i, idx;
487 
488 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
489 		if (adev->uvd.harvest_config & (1 << i))
490 			continue;
491 		if (adev->uvd.inst[i].vcpu_bo == NULL)
492 			return -EINVAL;
493 
494 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
495 		ptr = adev->uvd.inst[i].cpu_addr;
496 
497 		if (adev->uvd.inst[i].saved_bo != NULL) {
498 			if (drm_dev_enter(adev_to_drm(adev), &idx)) {
499 				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
500 				drm_dev_exit(idx);
501 			}
502 			kvfree(adev->uvd.inst[i].saved_bo);
503 			adev->uvd.inst[i].saved_bo = NULL;
504 		} else {
505 			const struct common_firmware_header *hdr;
506 			unsigned int offset;
507 
508 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
509 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
510 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
511 				if (drm_dev_enter(adev_to_drm(adev), &idx)) {
512 					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
513 						    le32_to_cpu(hdr->ucode_size_bytes));
514 					drm_dev_exit(idx);
515 				}
516 				size -= le32_to_cpu(hdr->ucode_size_bytes);
517 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
518 			}
519 			memset_io(ptr, 0, size);
520 			/* to restore uvd fence seq */
521 			if (adev->uvd.inst[i].ring.fence_drv.initialized)
522 				amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring, NULL);
523 		}
524 	}
525 	return 0;
526 }
527 
528 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
529 {
530 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
531 	int i, r;
532 
533 	for (i = 0; i < adev->uvd.max_handles; ++i) {
534 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
535 
536 		if (handle != 0 && adev->uvd.filp[i] == filp) {
537 			struct dma_fence *fence;
538 
539 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
540 						       &fence);
541 			if (r) {
542 				DRM_ERROR("Error destroying UVD %d!\n", r);
543 				continue;
544 			}
545 
546 			dma_fence_wait(fence, false);
547 			dma_fence_put(fence);
548 
549 			adev->uvd.filp[i] = NULL;
550 			atomic_set(&adev->uvd.handles[i], 0);
551 		}
552 	}
553 }
554 
555 static void amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo *bo)
556 {
557 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
558 	struct amdgpu_bo *vcpu_bo = adev->uvd.inst[0].vcpu_bo;
559 	struct amdgpu_res_cursor vcpu_cur;
560 
561 	amdgpu_res_first(vcpu_bo->tbo.resource, 0,
562 			 amdgpu_bo_size(vcpu_bo), &vcpu_cur);
563 
564 	bo->placement.num_placement = 1;
565 	bo->placement.placement = &bo->placements[0];
566 	bo->placements[0].fpfn = ALIGN_DOWN(vcpu_cur.start, SZ_256M) >> PAGE_SHIFT;
567 	bo->placements[0].lpfn = bo->placements[0].fpfn + (SZ_256M >> PAGE_SHIFT);
568 	bo->placements[0].mem_type = vcpu_bo->tbo.resource->mem_type;
569 	if (bo->placements[0].mem_type == TTM_PL_VRAM)
570 		bo->placements[0].flags |= TTM_PL_FLAG_CONTIGUOUS;
571 }
572 
573 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
574 {
575 	int i;
576 
577 	for (i = 0; i < abo->placement.num_placement; ++i) {
578 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
579 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
580 		if (abo->placements[i].mem_type == TTM_PL_VRAM)
581 			abo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
582 	}
583 }
584 
585 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
586 {
587 	uint32_t lo, hi;
588 	uint64_t addr;
589 
590 	lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
591 	hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
592 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
593 
594 	return addr;
595 }
596 
597 /**
598  * amdgpu_uvd_cs_pass1 - first parsing round
599  *
600  * @ctx: UVD parser context
601  *
602  * Make sure UVD message and feedback buffers are in VRAM and
603  * nobody is violating an 256MB boundary.
604  */
605 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
606 {
607 	struct ttm_operation_ctx tctx = { false, false };
608 	struct amdgpu_bo_va_mapping *mapping;
609 	struct amdgpu_bo *bo;
610 	uint32_t cmd;
611 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
612 	int r = 0;
613 
614 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
615 	if (r) {
616 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
617 		return r;
618 	}
619 
620 	if (!ctx->parser->adev->uvd.address_64_bit) {
621 		/* check if it's a message or feedback command */
622 		cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
623 		if (cmd == 0x0 || cmd == 0x3)
624 			amdgpu_uvd_force_into_vcpu_segment(bo);
625 		else
626 			amdgpu_uvd_force_into_uvd_segment(bo);
627 
628 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
629 	}
630 
631 	return r;
632 }
633 
634 /**
635  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
636  *
637  * @adev: amdgpu_device pointer
638  * @msg: pointer to message structure
639  * @buf_sizes: placeholder to put the different buffer lengths
640  *
641  * Peek into the decode message and calculate the necessary buffer sizes.
642  */
643 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
644 	unsigned int buf_sizes[])
645 {
646 	unsigned int stream_type = msg[4];
647 	unsigned int width = msg[6];
648 	unsigned int height = msg[7];
649 	unsigned int dpb_size = msg[9];
650 	unsigned int pitch = msg[28];
651 
652 	unsigned int width_in_mb = width / 16;
653 	unsigned int height_in_mb = ALIGN(height / 16, 2);
654 
655 	unsigned int image_size, tmp, min_dpb_size, num_dpb_buffer;
656 	unsigned int min_ctx_size = ~0;
657 
658 	/* Reject invalid dimensions */
659 	if (width < 16 || height < 16 || width > 4096 || height > 4096) {
660 		dev_WARN_ONCE(adev->dev, 1,
661 			      "Invalid UVD decoding dimensions (%dx%d)!\n",
662 			      width, height);
663 		return -EINVAL;
664 	}
665 
666 	image_size = width * height;
667 	image_size += image_size / 2;
668 	image_size = ALIGN(image_size, 1024);
669 
670 	switch (stream_type) {
671 	case 0: /* H264 */
672 		num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
673 		if (num_dpb_buffer > 17)
674 			return -EINVAL;
675 
676 		/* reference picture buffer */
677 		min_dpb_size = image_size * num_dpb_buffer;
678 
679 		/* macroblock context buffer */
680 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
681 
682 		/* IT surface buffer */
683 		min_dpb_size += width_in_mb * height_in_mb * 32;
684 		break;
685 
686 	case 1: /* VC1 */
687 
688 		/* reference picture buffer */
689 		min_dpb_size = image_size * 3;
690 
691 		/* CONTEXT_BUFFER */
692 		min_dpb_size += width_in_mb * height_in_mb * 128;
693 
694 		/* IT surface buffer */
695 		min_dpb_size += width_in_mb * 64;
696 
697 		/* DB surface buffer */
698 		min_dpb_size += width_in_mb * 128;
699 
700 		/* BP */
701 		tmp = max(width_in_mb, height_in_mb);
702 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
703 		break;
704 
705 	case 3: /* MPEG2 */
706 
707 		/* reference picture buffer */
708 		min_dpb_size = image_size * 3;
709 		break;
710 
711 	case 4: /* MPEG4 */
712 
713 		/* reference picture buffer */
714 		min_dpb_size = image_size * 3;
715 
716 		/* CM */
717 		min_dpb_size += width_in_mb * height_in_mb * 64;
718 
719 		/* IT surface buffer */
720 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
721 		break;
722 
723 	case 7: /* H264 Perf */
724 		num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
725 		if (num_dpb_buffer > 17)
726 			return -EINVAL;
727 
728 		/* reference picture buffer */
729 		min_dpb_size = image_size * num_dpb_buffer;
730 
731 		if (!adev->uvd.use_ctx_buf) {
732 			/* macroblock context buffer */
733 			min_dpb_size +=
734 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
735 
736 			/* IT surface buffer */
737 			min_dpb_size += width_in_mb * height_in_mb * 32;
738 		} else {
739 			/* macroblock context buffer */
740 			min_ctx_size =
741 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
742 		}
743 		break;
744 
745 	case 8: /* MJPEG */
746 		min_dpb_size = 0;
747 		break;
748 
749 	case 16: /* H265 */
750 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
751 		image_size = ALIGN(image_size, 256);
752 
753 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
754 		if (num_dpb_buffer > 17)
755 			return -EINVAL;
756 
757 		min_dpb_size = image_size * num_dpb_buffer;
758 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
759 					   * 16 * num_dpb_buffer + 52 * 1024;
760 		break;
761 
762 	default:
763 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
764 		return -EINVAL;
765 	}
766 
767 	if (width > pitch || pitch > 4096) {
768 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
769 		return -EINVAL;
770 	}
771 
772 	if (dpb_size < min_dpb_size) {
773 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
774 			  dpb_size, min_dpb_size);
775 		return -EINVAL;
776 	}
777 
778 	buf_sizes[0x1] = dpb_size;
779 	buf_sizes[0x2] = (pitch * height) * 3 / 2;
780 	buf_sizes[0x4] = min_ctx_size;
781 	/* store image width to adjust nb memory pstate */
782 	adev->uvd.decode_image_width = width;
783 	return 0;
784 }
785 
786 /**
787  * amdgpu_uvd_cs_msg - handle UVD message
788  *
789  * @ctx: UVD parser context
790  * @bo: buffer object containing the message
791  * @offset: offset into the buffer object
792  *
793  * Peek into the UVD message and extract the session id.
794  * Make sure that we don't open up to many sessions.
795  */
796 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
797 			     struct amdgpu_bo *bo, unsigned int offset)
798 {
799 	struct amdgpu_device *adev = ctx->parser->adev;
800 	int32_t *msg, msg_type, handle;
801 	void *ptr;
802 	long r;
803 	int i;
804 
805 	if (offset & 0x3F) {
806 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
807 		return -EINVAL;
808 	}
809 
810 	r = amdgpu_bo_kmap(bo, &ptr);
811 	if (r) {
812 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
813 		return r;
814 	}
815 
816 	msg = ptr + offset;
817 
818 	msg_type = msg[1];
819 	handle = msg[2];
820 
821 	if (handle == 0) {
822 		amdgpu_bo_kunmap(bo);
823 		DRM_ERROR("Invalid UVD handle!\n");
824 		return -EINVAL;
825 	}
826 
827 	switch (msg_type) {
828 	case 0:
829 		/* it's a create msg, calc image size (width * height) */
830 		amdgpu_bo_kunmap(bo);
831 
832 		/* try to alloc a new handle */
833 		for (i = 0; i < adev->uvd.max_handles; ++i) {
834 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
835 				DRM_ERROR(")Handle 0x%x already in use!\n",
836 					  handle);
837 				return -EINVAL;
838 			}
839 
840 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
841 				adev->uvd.filp[i] = ctx->parser->filp;
842 				return 0;
843 			}
844 		}
845 
846 		DRM_ERROR("No more free UVD handles!\n");
847 		return -ENOSPC;
848 
849 	case 1:
850 		/* it's a decode msg, calc buffer sizes */
851 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
852 		amdgpu_bo_kunmap(bo);
853 		if (r)
854 			return r;
855 
856 		/* validate the handle */
857 		for (i = 0; i < adev->uvd.max_handles; ++i) {
858 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
859 				if (adev->uvd.filp[i] != ctx->parser->filp) {
860 					DRM_ERROR("UVD handle collision detected!\n");
861 					return -EINVAL;
862 				}
863 				return 0;
864 			}
865 		}
866 
867 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
868 		return -ENOENT;
869 
870 	case 2:
871 		/* it's a destroy msg, free the handle */
872 		for (i = 0; i < adev->uvd.max_handles; ++i)
873 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
874 		amdgpu_bo_kunmap(bo);
875 		return 0;
876 
877 	default:
878 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
879 	}
880 
881 	amdgpu_bo_kunmap(bo);
882 	return -EINVAL;
883 }
884 
885 /**
886  * amdgpu_uvd_cs_pass2 - second parsing round
887  *
888  * @ctx: UVD parser context
889  *
890  * Patch buffer addresses, make sure buffer sizes are correct.
891  */
892 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
893 {
894 	struct amdgpu_bo_va_mapping *mapping;
895 	struct amdgpu_bo *bo;
896 	uint32_t cmd;
897 	uint64_t start, end;
898 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
899 	int r;
900 
901 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
902 	if (r) {
903 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
904 		return r;
905 	}
906 
907 	start = amdgpu_bo_gpu_offset(bo);
908 
909 	end = (mapping->last + 1 - mapping->start);
910 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
911 
912 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
913 	start += addr;
914 
915 	amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
916 	amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
917 
918 	cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
919 	if (cmd < 0x4) {
920 		if ((end - start) < ctx->buf_sizes[cmd]) {
921 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
922 				  (unsigned int)(end - start),
923 				  ctx->buf_sizes[cmd]);
924 			return -EINVAL;
925 		}
926 	} else if (cmd == 0x204 || cmd == 0x206) {
927 		unsigned int min_size = ctx->buf_sizes[cmd == 0x204 ? 5 : 4];
928 
929 		if ((end - start) < min_size) {
930 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
931 					  (unsigned int)(end - start),
932 					  min_size);
933 			return -EINVAL;
934 		}
935 	} else if ((cmd != 0x100)) {
936 		DRM_ERROR("invalid UVD command %X!\n", cmd);
937 		return -EINVAL;
938 	}
939 
940 	if (!ctx->parser->adev->uvd.address_64_bit) {
941 		if ((start >> 28) != ((end - 1) >> 28)) {
942 			DRM_ERROR("reloc %llx-%llx crossing 256MB boundary!\n",
943 				  start, end);
944 			return -EINVAL;
945 		}
946 
947 		if ((cmd == 0 || cmd == 0x3) &&
948 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
949 			DRM_ERROR("msg/fb buffer %llx-%llx out of 256MB segment!\n",
950 				  start, end);
951 			return -EINVAL;
952 		}
953 	}
954 
955 	if (cmd == 0) {
956 		ctx->has_msg_cmd = true;
957 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
958 		if (r)
959 			return r;
960 	} else if (!ctx->has_msg_cmd) {
961 		DRM_ERROR("Message needed before other commands are send!\n");
962 		return -EINVAL;
963 	}
964 
965 	return 0;
966 }
967 
968 /**
969  * amdgpu_uvd_cs_reg - parse register writes
970  *
971  * @ctx: UVD parser context
972  * @cb: callback function
973  *
974  * Parse the register writes, call cb on each complete command.
975  */
976 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
977 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
978 {
979 	int i, r;
980 
981 	ctx->idx++;
982 	for (i = 0; i <= ctx->count; ++i) {
983 		unsigned int reg = ctx->reg + i;
984 
985 		if (ctx->idx >= ctx->ib->length_dw) {
986 			DRM_ERROR("Register command after end of CS!\n");
987 			return -EINVAL;
988 		}
989 
990 		switch (reg) {
991 		case mmUVD_GPCOM_VCPU_DATA0:
992 			ctx->data0 = ctx->idx;
993 			break;
994 		case mmUVD_GPCOM_VCPU_DATA1:
995 			ctx->data1 = ctx->idx;
996 			break;
997 		case mmUVD_GPCOM_VCPU_CMD:
998 			r = cb(ctx);
999 			if (r)
1000 				return r;
1001 			break;
1002 		case mmUVD_ENGINE_CNTL:
1003 		case mmUVD_NO_OP:
1004 			break;
1005 		default:
1006 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1007 			return -EINVAL;
1008 		}
1009 		ctx->idx++;
1010 	}
1011 	return 0;
1012 }
1013 
1014 /**
1015  * amdgpu_uvd_cs_packets - parse UVD packets
1016  *
1017  * @ctx: UVD parser context
1018  * @cb: callback function
1019  *
1020  * Parse the command stream packets.
1021  */
1022 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1023 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1024 {
1025 	int r;
1026 
1027 	for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1028 		uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1029 		unsigned int type = CP_PACKET_GET_TYPE(cmd);
1030 
1031 		switch (type) {
1032 		case PACKET_TYPE0:
1033 			ctx->reg = CP_PACKET0_GET_REG(cmd);
1034 			ctx->count = CP_PACKET_GET_COUNT(cmd);
1035 			r = amdgpu_uvd_cs_reg(ctx, cb);
1036 			if (r)
1037 				return r;
1038 			break;
1039 		case PACKET_TYPE2:
1040 			++ctx->idx;
1041 			break;
1042 		default:
1043 			DRM_ERROR("Unknown packet type %d !\n", type);
1044 			return -EINVAL;
1045 		}
1046 	}
1047 	return 0;
1048 }
1049 
1050 /**
1051  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1052  *
1053  * @parser: Command submission parser context
1054  * @job: the job to parse
1055  * @ib: the IB to patch
1056  *
1057  * Parse the command stream, patch in addresses as necessary.
1058  */
1059 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1060 			     struct amdgpu_job *job,
1061 			     struct amdgpu_ib *ib)
1062 {
1063 	struct amdgpu_uvd_cs_ctx ctx = {};
1064 	unsigned int buf_sizes[] = {
1065 		[0x00000000]	=	3556,
1066 		[0x00000001]	=	0xFFFFFFFF,
1067 		[0x00000002]	=	0xFFFFFFFF,
1068 		[0x00000003]	=	2048,
1069 		[0x00000004]	=	0xFFFFFFFF,
1070 		[0x00000005]	=	992,
1071 	};
1072 	int r;
1073 
1074 	job->vm = NULL;
1075 
1076 	if (ib->length_dw % 16) {
1077 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1078 			  ib->length_dw);
1079 		return -EINVAL;
1080 	}
1081 
1082 	ctx.parser = parser;
1083 	ctx.buf_sizes = buf_sizes;
1084 	ctx.ib = ib;
1085 
1086 	/* first round only required on chips without UVD 64 bit address support */
1087 	if (!parser->adev->uvd.address_64_bit) {
1088 		/* first round, make sure the buffers are actually in the UVD segment */
1089 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1090 		if (r)
1091 			return r;
1092 	}
1093 
1094 	/* second round, patch buffer addresses into the command stream */
1095 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1096 	if (r)
1097 		return r;
1098 
1099 	if (!ctx.has_msg_cmd) {
1100 		DRM_ERROR("UVD-IBs need a msg command!\n");
1101 		return -EINVAL;
1102 	}
1103 
1104 	return 0;
1105 }
1106 
1107 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1108 			       bool direct, struct dma_fence **fence)
1109 {
1110 	struct amdgpu_device *adev = ring->adev;
1111 	struct dma_fence *f = NULL;
1112 	uint32_t offset, data[4];
1113 	struct amdgpu_job *job;
1114 	struct amdgpu_ib *ib;
1115 	uint64_t addr;
1116 	int i, r;
1117 
1118 	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->uvd.entity,
1119 				     AMDGPU_FENCE_OWNER_UNDEFINED,
1120 				     64, direct ? AMDGPU_IB_POOL_DIRECT :
1121 				     AMDGPU_IB_POOL_DELAYED,
1122 				     AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST,
1123 				     &job);
1124 	if (r)
1125 		return r;
1126 
1127 	if (adev->asic_type >= CHIP_VEGA10)
1128 		offset = adev->reg_offset[UVD_HWIP][ring->me][1];
1129 	else
1130 		offset = UVD_BASE_SI;
1131 
1132 	data[0] = PACKET0(offset + UVD_GPCOM_VCPU_DATA0, 0);
1133 	data[1] = PACKET0(offset + UVD_GPCOM_VCPU_DATA1, 0);
1134 	data[2] = PACKET0(offset + UVD_GPCOM_VCPU_CMD, 0);
1135 	data[3] = PACKET0(offset + UVD_NO_OP, 0);
1136 
1137 	ib = &job->ibs[0];
1138 	addr = amdgpu_bo_gpu_offset(bo);
1139 	ib->ptr[0] = data[0];
1140 	ib->ptr[1] = addr;
1141 	ib->ptr[2] = data[1];
1142 	ib->ptr[3] = addr >> 32;
1143 	ib->ptr[4] = data[2];
1144 	ib->ptr[5] = 0;
1145 	for (i = 6; i < 16; i += 2) {
1146 		ib->ptr[i] = data[3];
1147 		ib->ptr[i+1] = 0;
1148 	}
1149 	ib->length_dw = 16;
1150 
1151 	if (direct) {
1152 		r = amdgpu_job_submit_direct(job, ring, &f);
1153 		if (r)
1154 			goto err_free;
1155 	} else {
1156 		r = drm_sched_job_add_resv_dependencies(&job->base,
1157 							bo->tbo.base.resv,
1158 							DMA_RESV_USAGE_KERNEL);
1159 		if (r)
1160 			goto err_free;
1161 
1162 		f = amdgpu_job_submit(job);
1163 	}
1164 
1165 	amdgpu_bo_reserve(bo, true);
1166 	amdgpu_bo_fence(bo, f, false);
1167 	amdgpu_bo_unreserve(bo);
1168 
1169 	if (fence)
1170 		*fence = dma_fence_get(f);
1171 	dma_fence_put(f);
1172 
1173 	return 0;
1174 
1175 err_free:
1176 	amdgpu_job_free(job);
1177 	return r;
1178 }
1179 
1180 /* multiple fence commands without any stream commands in between can
1181  * crash the vcpu so just try to emmit a dummy create/destroy msg to
1182  * avoid this
1183  */
1184 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1185 			      struct dma_fence **fence)
1186 {
1187 	struct amdgpu_device *adev = ring->adev;
1188 	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1189 	uint32_t *msg;
1190 	int i;
1191 
1192 	msg = amdgpu_bo_kptr(bo);
1193 	/* stitch together an UVD create msg */
1194 	msg[0] = cpu_to_le32(0x00000de4);
1195 	msg[1] = cpu_to_le32(0x00000000);
1196 	msg[2] = cpu_to_le32(handle);
1197 	msg[3] = cpu_to_le32(0x00000000);
1198 	msg[4] = cpu_to_le32(0x00000000);
1199 	msg[5] = cpu_to_le32(0x00000000);
1200 	msg[6] = cpu_to_le32(0x00000000);
1201 	msg[7] = cpu_to_le32(0x00000780);
1202 	msg[8] = cpu_to_le32(0x00000440);
1203 	msg[9] = cpu_to_le32(0x00000000);
1204 	msg[10] = cpu_to_le32(0x01b37000);
1205 	for (i = 11; i < 1024; ++i)
1206 		msg[i] = cpu_to_le32(0x0);
1207 
1208 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1209 
1210 }
1211 
1212 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1213 			       bool direct, struct dma_fence **fence)
1214 {
1215 	struct amdgpu_device *adev = ring->adev;
1216 	struct amdgpu_bo *bo = NULL;
1217 	uint32_t *msg;
1218 	int r, i;
1219 
1220 	if (direct) {
1221 		bo = adev->uvd.ib_bo;
1222 	} else {
1223 		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1224 		if (r)
1225 			return r;
1226 	}
1227 
1228 	msg = amdgpu_bo_kptr(bo);
1229 	/* stitch together an UVD destroy msg */
1230 	msg[0] = cpu_to_le32(0x00000de4);
1231 	msg[1] = cpu_to_le32(0x00000002);
1232 	msg[2] = cpu_to_le32(handle);
1233 	msg[3] = cpu_to_le32(0x00000000);
1234 	for (i = 4; i < 1024; ++i)
1235 		msg[i] = cpu_to_le32(0x0);
1236 
1237 	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1238 
1239 	if (!direct)
1240 		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1241 
1242 	return r;
1243 }
1244 
1245 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1246 {
1247 	struct amdgpu_device *adev =
1248 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1249 	unsigned int fences = 0, i, j;
1250 
1251 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1252 		if (adev->uvd.harvest_config & (1 << i))
1253 			continue;
1254 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1255 		for (j = 0; j < adev->uvd.num_enc_rings; ++j)
1256 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1257 	}
1258 
1259 	if (fences == 0) {
1260 		if (adev->pm.dpm_enabled) {
1261 			amdgpu_dpm_enable_uvd(adev, false);
1262 		} else {
1263 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1264 			/* shutdown the UVD block */
1265 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1266 							       AMD_PG_STATE_GATE);
1267 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1268 							       AMD_CG_STATE_GATE);
1269 		}
1270 	} else {
1271 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1272 	}
1273 }
1274 
1275 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1276 {
1277 	struct amdgpu_device *adev = ring->adev;
1278 	bool set_clocks;
1279 
1280 	if (amdgpu_sriov_vf(adev))
1281 		return;
1282 
1283 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1284 	if (set_clocks) {
1285 		if (adev->pm.dpm_enabled) {
1286 			amdgpu_dpm_enable_uvd(adev, true);
1287 		} else {
1288 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1289 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1290 							       AMD_CG_STATE_UNGATE);
1291 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1292 							       AMD_PG_STATE_UNGATE);
1293 		}
1294 	}
1295 }
1296 
1297 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1298 {
1299 	if (!amdgpu_sriov_vf(ring->adev))
1300 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1301 }
1302 
1303 /**
1304  * amdgpu_uvd_ring_test_ib - test ib execution
1305  *
1306  * @ring: amdgpu_ring pointer
1307  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1308  *
1309  * Test if we can successfully execute an IB
1310  */
1311 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1312 {
1313 	struct dma_fence *fence;
1314 	long r;
1315 
1316 	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1317 	if (r)
1318 		goto error;
1319 
1320 	r = dma_fence_wait_timeout(fence, false, timeout);
1321 	dma_fence_put(fence);
1322 	if (r == 0)
1323 		r = -ETIMEDOUT;
1324 	if (r < 0)
1325 		goto error;
1326 
1327 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1328 	if (r)
1329 		goto error;
1330 
1331 	r = dma_fence_wait_timeout(fence, false, timeout);
1332 	if (r == 0)
1333 		r = -ETIMEDOUT;
1334 	else if (r > 0)
1335 		r = 0;
1336 
1337 	dma_fence_put(fence);
1338 
1339 error:
1340 	return r;
1341 }
1342 
1343 /**
1344  * amdgpu_uvd_used_handles - returns used UVD handles
1345  *
1346  * @adev: amdgpu_device pointer
1347  *
1348  * Returns the number of UVD handles in use
1349  */
1350 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1351 {
1352 	unsigned int i;
1353 	uint32_t used_handles = 0;
1354 
1355 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1356 		/*
1357 		 * Handles can be freed in any order, and not
1358 		 * necessarily linear. So we need to count
1359 		 * all non-zero handles.
1360 		 */
1361 		if (atomic_read(&adev->uvd.handles[i]))
1362 			used_handles++;
1363 	}
1364 
1365 	return used_handles;
1366 }
1367