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 reg, count;
100 	unsigned data0, data1;
101 	unsigned 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 *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_uvd_segment(struct amdgpu_bo *abo);
139 
amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device * adev,uint32_t size,struct amdgpu_bo ** bo_ptr)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_uvd_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 
amdgpu_uvd_sw_init(struct amdgpu_device * adev)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 family_id;
190 	int i, j, r;
191 
192 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
193 
194 	switch (adev->asic_type) {
195 #ifdef CONFIG_DRM_AMDGPU_SI
196 	case CHIP_TAHITI:
197 		fw_name = FIRMWARE_TAHITI;
198 		break;
199 	case CHIP_VERDE:
200 		fw_name = FIRMWARE_VERDE;
201 		break;
202 	case CHIP_PITCAIRN:
203 		fw_name = FIRMWARE_PITCAIRN;
204 		break;
205 	case CHIP_OLAND:
206 		fw_name = FIRMWARE_OLAND;
207 		break;
208 #endif
209 #ifdef CONFIG_DRM_AMDGPU_CIK
210 	case CHIP_BONAIRE:
211 		fw_name = FIRMWARE_BONAIRE;
212 		break;
213 	case CHIP_KABINI:
214 		fw_name = FIRMWARE_KABINI;
215 		break;
216 	case CHIP_KAVERI:
217 		fw_name = FIRMWARE_KAVERI;
218 		break;
219 	case CHIP_HAWAII:
220 		fw_name = FIRMWARE_HAWAII;
221 		break;
222 	case CHIP_MULLINS:
223 		fw_name = FIRMWARE_MULLINS;
224 		break;
225 #endif
226 	case CHIP_TONGA:
227 		fw_name = FIRMWARE_TONGA;
228 		break;
229 	case CHIP_FIJI:
230 		fw_name = FIRMWARE_FIJI;
231 		break;
232 	case CHIP_CARRIZO:
233 		fw_name = FIRMWARE_CARRIZO;
234 		break;
235 	case CHIP_STONEY:
236 		fw_name = FIRMWARE_STONEY;
237 		break;
238 	case CHIP_POLARIS10:
239 		fw_name = FIRMWARE_POLARIS10;
240 		break;
241 	case CHIP_POLARIS11:
242 		fw_name = FIRMWARE_POLARIS11;
243 		break;
244 	case CHIP_POLARIS12:
245 		fw_name = FIRMWARE_POLARIS12;
246 		break;
247 	case CHIP_VEGA10:
248 		fw_name = FIRMWARE_VEGA10;
249 		break;
250 	case CHIP_VEGA12:
251 		fw_name = FIRMWARE_VEGA12;
252 		break;
253 	case CHIP_VEGAM:
254 		fw_name = FIRMWARE_VEGAM;
255 		break;
256 	case CHIP_VEGA20:
257 		fw_name = FIRMWARE_VEGA20;
258 		break;
259 	default:
260 		return -EINVAL;
261 	}
262 
263 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
264 	if (r) {
265 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
266 			fw_name);
267 		return r;
268 	}
269 
270 	r = amdgpu_ucode_validate(adev->uvd.fw);
271 	if (r) {
272 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
273 			fw_name);
274 		release_firmware(adev->uvd.fw);
275 		adev->uvd.fw = NULL;
276 		return r;
277 	}
278 
279 	/* Set the default UVD handles that the firmware can handle */
280 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
281 
282 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
283 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
284 
285 	if (adev->asic_type < CHIP_VEGA20) {
286 		unsigned version_major, version_minor;
287 
288 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
289 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
290 		DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
291 			version_major, version_minor, family_id);
292 
293 		/*
294 		 * Limit the number of UVD handles depending on microcode major
295 		 * and minor versions. The firmware version which has 40 UVD
296 		 * instances support is 1.80. So all subsequent versions should
297 		 * also have the same support.
298 		 */
299 		if ((version_major > 0x01) ||
300 		    ((version_major == 0x01) && (version_minor >= 0x50)))
301 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
302 
303 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
304 					(family_id << 8));
305 
306 		if ((adev->asic_type == CHIP_POLARIS10 ||
307 		     adev->asic_type == CHIP_POLARIS11) &&
308 		    (adev->uvd.fw_version < FW_1_66_16))
309 			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
310 				  version_major, version_minor);
311 	} else {
312 		unsigned int enc_major, enc_minor, dec_minor;
313 
314 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
315 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
316 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
317 		DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
318 			enc_major, enc_minor, dec_minor, family_id);
319 
320 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
321 
322 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
323 	}
324 
325 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
326 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
327 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
328 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
329 
330 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
331 		if (adev->uvd.harvest_config & (1 << j))
332 			continue;
333 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
334 					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
335 					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
336 		if (r) {
337 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
338 			return r;
339 		}
340 	}
341 
342 	for (i = 0; i < adev->uvd.max_handles; ++i) {
343 		atomic_set(&adev->uvd.handles[i], 0);
344 		adev->uvd.filp[i] = NULL;
345 	}
346 
347 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
348 	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
349 		adev->uvd.address_64_bit = true;
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 
amdgpu_uvd_sw_fini(struct amdgpu_device * adev)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 	drm_sched_entity_destroy(&adev->uvd.entity);
381 
382 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
383 		if (adev->uvd.harvest_config & (1 << j))
384 			continue;
385 		kvfree(adev->uvd.inst[j].saved_bo);
386 
387 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
388 				      &adev->uvd.inst[j].gpu_addr,
389 				      (void **)&adev->uvd.inst[j].cpu_addr);
390 
391 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
392 
393 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
394 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
395 	}
396 	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
397 	release_firmware(adev->uvd.fw);
398 
399 	return 0;
400 }
401 
402 /**
403  * amdgpu_uvd_entity_init - init entity
404  *
405  * @adev: amdgpu_device pointer
406  *
407  */
amdgpu_uvd_entity_init(struct amdgpu_device * adev)408 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
409 {
410 	struct amdgpu_ring *ring;
411 	struct drm_gpu_scheduler *sched;
412 	int r;
413 
414 	ring = &adev->uvd.inst[0].ring;
415 	sched = &ring->sched;
416 	r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
417 				  &sched, 1, NULL);
418 	if (r) {
419 		DRM_ERROR("Failed setting up UVD kernel entity.\n");
420 		return r;
421 	}
422 
423 	return 0;
424 }
425 
amdgpu_uvd_suspend(struct amdgpu_device * adev)426 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
427 {
428 	unsigned size;
429 	void *ptr;
430 	int i, j, idx;
431 	bool in_ras_intr = amdgpu_ras_intr_triggered();
432 
433 	cancel_delayed_work_sync(&adev->uvd.idle_work);
434 
435 	/* only valid for physical mode */
436 	if (adev->asic_type < CHIP_POLARIS10) {
437 		for (i = 0; i < adev->uvd.max_handles; ++i)
438 			if (atomic_read(&adev->uvd.handles[i]))
439 				break;
440 
441 		if (i == adev->uvd.max_handles)
442 			return 0;
443 	}
444 
445 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
446 		if (adev->uvd.harvest_config & (1 << j))
447 			continue;
448 		if (adev->uvd.inst[j].vcpu_bo == NULL)
449 			continue;
450 
451 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
452 		ptr = adev->uvd.inst[j].cpu_addr;
453 
454 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
455 		if (!adev->uvd.inst[j].saved_bo)
456 			return -ENOMEM;
457 
458 		if (drm_dev_enter(adev_to_drm(adev), &idx)) {
459 			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
460 			if (in_ras_intr)
461 				memset(adev->uvd.inst[j].saved_bo, 0, size);
462 			else
463 				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
464 
465 			drm_dev_exit(idx);
466 		}
467 	}
468 
469 	if (in_ras_intr)
470 		DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
471 
472 	return 0;
473 }
474 
amdgpu_uvd_resume(struct amdgpu_device * adev)475 int amdgpu_uvd_resume(struct amdgpu_device *adev)
476 {
477 	unsigned size;
478 	void *ptr;
479 	int i, idx;
480 
481 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
482 		if (adev->uvd.harvest_config & (1 << i))
483 			continue;
484 		if (adev->uvd.inst[i].vcpu_bo == NULL)
485 			return -EINVAL;
486 
487 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
488 		ptr = adev->uvd.inst[i].cpu_addr;
489 
490 		if (adev->uvd.inst[i].saved_bo != NULL) {
491 			if (drm_dev_enter(adev_to_drm(adev), &idx)) {
492 				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
493 				drm_dev_exit(idx);
494 			}
495 			kvfree(adev->uvd.inst[i].saved_bo);
496 			adev->uvd.inst[i].saved_bo = NULL;
497 		} else {
498 			const struct common_firmware_header *hdr;
499 			unsigned offset;
500 
501 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
502 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
503 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
504 				if (drm_dev_enter(adev_to_drm(adev), &idx)) {
505 					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
506 						    le32_to_cpu(hdr->ucode_size_bytes));
507 					drm_dev_exit(idx);
508 				}
509 				size -= le32_to_cpu(hdr->ucode_size_bytes);
510 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
511 			}
512 			memset_io(ptr, 0, size);
513 			/* to restore uvd fence seq */
514 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
515 		}
516 	}
517 	return 0;
518 }
519 
amdgpu_uvd_free_handles(struct amdgpu_device * adev,struct drm_file * filp)520 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
521 {
522 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
523 	int i, r;
524 
525 	for (i = 0; i < adev->uvd.max_handles; ++i) {
526 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
527 
528 		if (handle != 0 && adev->uvd.filp[i] == filp) {
529 			struct dma_fence *fence;
530 
531 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
532 						       &fence);
533 			if (r) {
534 				DRM_ERROR("Error destroying UVD %d!\n", r);
535 				continue;
536 			}
537 
538 			dma_fence_wait(fence, false);
539 			dma_fence_put(fence);
540 
541 			adev->uvd.filp[i] = NULL;
542 			atomic_set(&adev->uvd.handles[i], 0);
543 		}
544 	}
545 }
546 
amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo * abo)547 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
548 {
549 	int i;
550 	for (i = 0; i < abo->placement.num_placement; ++i) {
551 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
552 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
553 	}
554 }
555 
amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx * ctx)556 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
557 {
558 	uint32_t lo, hi;
559 	uint64_t addr;
560 
561 	lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
562 	hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
563 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
564 
565 	return addr;
566 }
567 
568 /**
569  * amdgpu_uvd_cs_pass1 - first parsing round
570  *
571  * @ctx: UVD parser context
572  *
573  * Make sure UVD message and feedback buffers are in VRAM and
574  * nobody is violating an 256MB boundary.
575  */
amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx * ctx)576 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
577 {
578 	struct ttm_operation_ctx tctx = { false, false };
579 	struct amdgpu_bo_va_mapping *mapping;
580 	struct amdgpu_bo *bo;
581 	uint32_t cmd;
582 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
583 	int r = 0;
584 
585 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
586 	if (r) {
587 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
588 		return r;
589 	}
590 
591 	if (!ctx->parser->adev->uvd.address_64_bit) {
592 		/* check if it's a message or feedback command */
593 		cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
594 		if (cmd == 0x0 || cmd == 0x3) {
595 			/* yes, force it into VRAM */
596 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
597 			amdgpu_bo_placement_from_domain(bo, domain);
598 		}
599 		amdgpu_uvd_force_into_uvd_segment(bo);
600 
601 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
602 	}
603 
604 	return r;
605 }
606 
607 /**
608  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
609  *
610  * @adev: amdgpu_device pointer
611  * @msg: pointer to message structure
612  * @buf_sizes: placeholder to put the different buffer lengths
613  *
614  * Peek into the decode message and calculate the necessary buffer sizes.
615  */
amdgpu_uvd_cs_msg_decode(struct amdgpu_device * adev,uint32_t * msg,unsigned buf_sizes[])616 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
617 	unsigned buf_sizes[])
618 {
619 	unsigned stream_type = msg[4];
620 	unsigned width = msg[6];
621 	unsigned height = msg[7];
622 	unsigned dpb_size = msg[9];
623 	unsigned pitch = msg[28];
624 	unsigned level = msg[57];
625 
626 	unsigned width_in_mb = width / 16;
627 	unsigned height_in_mb = ALIGN(height / 16, 2);
628 	unsigned fs_in_mb = width_in_mb * height_in_mb;
629 
630 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
631 	unsigned min_ctx_size = ~0;
632 
633 	image_size = width * height;
634 	image_size += image_size / 2;
635 	image_size = ALIGN(image_size, 1024);
636 
637 	switch (stream_type) {
638 	case 0: /* H264 */
639 		switch(level) {
640 		case 30:
641 			num_dpb_buffer = 8100 / fs_in_mb;
642 			break;
643 		case 31:
644 			num_dpb_buffer = 18000 / fs_in_mb;
645 			break;
646 		case 32:
647 			num_dpb_buffer = 20480 / fs_in_mb;
648 			break;
649 		case 41:
650 			num_dpb_buffer = 32768 / fs_in_mb;
651 			break;
652 		case 42:
653 			num_dpb_buffer = 34816 / fs_in_mb;
654 			break;
655 		case 50:
656 			num_dpb_buffer = 110400 / fs_in_mb;
657 			break;
658 		case 51:
659 			num_dpb_buffer = 184320 / fs_in_mb;
660 			break;
661 		default:
662 			num_dpb_buffer = 184320 / fs_in_mb;
663 			break;
664 		}
665 		num_dpb_buffer++;
666 		if (num_dpb_buffer > 17)
667 			num_dpb_buffer = 17;
668 
669 		/* reference picture buffer */
670 		min_dpb_size = image_size * num_dpb_buffer;
671 
672 		/* macroblock context buffer */
673 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
674 
675 		/* IT surface buffer */
676 		min_dpb_size += width_in_mb * height_in_mb * 32;
677 		break;
678 
679 	case 1: /* VC1 */
680 
681 		/* reference picture buffer */
682 		min_dpb_size = image_size * 3;
683 
684 		/* CONTEXT_BUFFER */
685 		min_dpb_size += width_in_mb * height_in_mb * 128;
686 
687 		/* IT surface buffer */
688 		min_dpb_size += width_in_mb * 64;
689 
690 		/* DB surface buffer */
691 		min_dpb_size += width_in_mb * 128;
692 
693 		/* BP */
694 		tmp = max(width_in_mb, height_in_mb);
695 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
696 		break;
697 
698 	case 3: /* MPEG2 */
699 
700 		/* reference picture buffer */
701 		min_dpb_size = image_size * 3;
702 		break;
703 
704 	case 4: /* MPEG4 */
705 
706 		/* reference picture buffer */
707 		min_dpb_size = image_size * 3;
708 
709 		/* CM */
710 		min_dpb_size += width_in_mb * height_in_mb * 64;
711 
712 		/* IT surface buffer */
713 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
714 		break;
715 
716 	case 7: /* H264 Perf */
717 		switch(level) {
718 		case 30:
719 			num_dpb_buffer = 8100 / fs_in_mb;
720 			break;
721 		case 31:
722 			num_dpb_buffer = 18000 / fs_in_mb;
723 			break;
724 		case 32:
725 			num_dpb_buffer = 20480 / fs_in_mb;
726 			break;
727 		case 41:
728 			num_dpb_buffer = 32768 / fs_in_mb;
729 			break;
730 		case 42:
731 			num_dpb_buffer = 34816 / fs_in_mb;
732 			break;
733 		case 50:
734 			num_dpb_buffer = 110400 / fs_in_mb;
735 			break;
736 		case 51:
737 			num_dpb_buffer = 184320 / fs_in_mb;
738 			break;
739 		default:
740 			num_dpb_buffer = 184320 / fs_in_mb;
741 			break;
742 		}
743 		num_dpb_buffer++;
744 		if (num_dpb_buffer > 17)
745 			num_dpb_buffer = 17;
746 
747 		/* reference picture buffer */
748 		min_dpb_size = image_size * num_dpb_buffer;
749 
750 		if (!adev->uvd.use_ctx_buf){
751 			/* macroblock context buffer */
752 			min_dpb_size +=
753 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
754 
755 			/* IT surface buffer */
756 			min_dpb_size += width_in_mb * height_in_mb * 32;
757 		} else {
758 			/* macroblock context buffer */
759 			min_ctx_size =
760 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
761 		}
762 		break;
763 
764 	case 8: /* MJPEG */
765 		min_dpb_size = 0;
766 		break;
767 
768 	case 16: /* H265 */
769 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
770 		image_size = ALIGN(image_size, 256);
771 
772 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
773 		min_dpb_size = image_size * num_dpb_buffer;
774 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
775 					   * 16 * num_dpb_buffer + 52 * 1024;
776 		break;
777 
778 	default:
779 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
780 		return -EINVAL;
781 	}
782 
783 	if (width > pitch) {
784 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
785 		return -EINVAL;
786 	}
787 
788 	if (dpb_size < min_dpb_size) {
789 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
790 			  dpb_size, min_dpb_size);
791 		return -EINVAL;
792 	}
793 
794 	buf_sizes[0x1] = dpb_size;
795 	buf_sizes[0x2] = image_size;
796 	buf_sizes[0x4] = min_ctx_size;
797 	/* store image width to adjust nb memory pstate */
798 	adev->uvd.decode_image_width = width;
799 	return 0;
800 }
801 
802 /**
803  * amdgpu_uvd_cs_msg - handle UVD message
804  *
805  * @ctx: UVD parser context
806  * @bo: buffer object containing the message
807  * @offset: offset into the buffer object
808  *
809  * Peek into the UVD message and extract the session id.
810  * Make sure that we don't open up to many sessions.
811  */
amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx * ctx,struct amdgpu_bo * bo,unsigned offset)812 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
813 			     struct amdgpu_bo *bo, unsigned offset)
814 {
815 	struct amdgpu_device *adev = ctx->parser->adev;
816 	int32_t *msg, msg_type, handle;
817 	void *ptr;
818 	long r;
819 	int i;
820 
821 	if (offset & 0x3F) {
822 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
823 		return -EINVAL;
824 	}
825 
826 	r = amdgpu_bo_kmap(bo, &ptr);
827 	if (r) {
828 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
829 		return r;
830 	}
831 
832 	msg = ptr + offset;
833 
834 	msg_type = msg[1];
835 	handle = msg[2];
836 
837 	if (handle == 0) {
838 		amdgpu_bo_kunmap(bo);
839 		DRM_ERROR("Invalid UVD handle!\n");
840 		return -EINVAL;
841 	}
842 
843 	switch (msg_type) {
844 	case 0:
845 		/* it's a create msg, calc image size (width * height) */
846 		amdgpu_bo_kunmap(bo);
847 
848 		/* try to alloc a new handle */
849 		for (i = 0; i < adev->uvd.max_handles; ++i) {
850 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
851 				DRM_ERROR(")Handle 0x%x already in use!\n",
852 					  handle);
853 				return -EINVAL;
854 			}
855 
856 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
857 				adev->uvd.filp[i] = ctx->parser->filp;
858 				return 0;
859 			}
860 		}
861 
862 		DRM_ERROR("No more free UVD handles!\n");
863 		return -ENOSPC;
864 
865 	case 1:
866 		/* it's a decode msg, calc buffer sizes */
867 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
868 		amdgpu_bo_kunmap(bo);
869 		if (r)
870 			return r;
871 
872 		/* validate the handle */
873 		for (i = 0; i < adev->uvd.max_handles; ++i) {
874 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
875 				if (adev->uvd.filp[i] != ctx->parser->filp) {
876 					DRM_ERROR("UVD handle collision detected!\n");
877 					return -EINVAL;
878 				}
879 				return 0;
880 			}
881 		}
882 
883 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
884 		return -ENOENT;
885 
886 	case 2:
887 		/* it's a destroy msg, free the handle */
888 		for (i = 0; i < adev->uvd.max_handles; ++i)
889 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
890 		amdgpu_bo_kunmap(bo);
891 		return 0;
892 
893 	default:
894 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
895 	}
896 
897 	amdgpu_bo_kunmap(bo);
898 	return -EINVAL;
899 }
900 
901 /**
902  * amdgpu_uvd_cs_pass2 - second parsing round
903  *
904  * @ctx: UVD parser context
905  *
906  * Patch buffer addresses, make sure buffer sizes are correct.
907  */
amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx * ctx)908 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
909 {
910 	struct amdgpu_bo_va_mapping *mapping;
911 	struct amdgpu_bo *bo;
912 	uint32_t cmd;
913 	uint64_t start, end;
914 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
915 	int r;
916 
917 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
918 	if (r) {
919 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
920 		return r;
921 	}
922 
923 	start = amdgpu_bo_gpu_offset(bo);
924 
925 	end = (mapping->last + 1 - mapping->start);
926 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
927 
928 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
929 	start += addr;
930 
931 	amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
932 	amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
933 
934 	cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
935 	if (cmd < 0x4) {
936 		if ((end - start) < ctx->buf_sizes[cmd]) {
937 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
938 				  (unsigned)(end - start),
939 				  ctx->buf_sizes[cmd]);
940 			return -EINVAL;
941 		}
942 
943 	} else if (cmd == 0x206) {
944 		if ((end - start) < ctx->buf_sizes[4]) {
945 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
946 					  (unsigned)(end - start),
947 					  ctx->buf_sizes[4]);
948 			return -EINVAL;
949 		}
950 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
951 		DRM_ERROR("invalid UVD command %X!\n", cmd);
952 		return -EINVAL;
953 	}
954 
955 	if (!ctx->parser->adev->uvd.address_64_bit) {
956 		if ((start >> 28) != ((end - 1) >> 28)) {
957 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
958 				  start, end);
959 			return -EINVAL;
960 		}
961 
962 		if ((cmd == 0 || cmd == 0x3) &&
963 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
964 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
965 				  start, end);
966 			return -EINVAL;
967 		}
968 	}
969 
970 	if (cmd == 0) {
971 		ctx->has_msg_cmd = true;
972 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
973 		if (r)
974 			return r;
975 	} else if (!ctx->has_msg_cmd) {
976 		DRM_ERROR("Message needed before other commands are send!\n");
977 		return -EINVAL;
978 	}
979 
980 	return 0;
981 }
982 
983 /**
984  * amdgpu_uvd_cs_reg - parse register writes
985  *
986  * @ctx: UVD parser context
987  * @cb: callback function
988  *
989  * Parse the register writes, call cb on each complete command.
990  */
amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))991 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
992 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
993 {
994 	int i, r;
995 
996 	ctx->idx++;
997 	for (i = 0; i <= ctx->count; ++i) {
998 		unsigned reg = ctx->reg + i;
999 
1000 		if (ctx->idx >= ctx->ib->length_dw) {
1001 			DRM_ERROR("Register command after end of CS!\n");
1002 			return -EINVAL;
1003 		}
1004 
1005 		switch (reg) {
1006 		case mmUVD_GPCOM_VCPU_DATA0:
1007 			ctx->data0 = ctx->idx;
1008 			break;
1009 		case mmUVD_GPCOM_VCPU_DATA1:
1010 			ctx->data1 = ctx->idx;
1011 			break;
1012 		case mmUVD_GPCOM_VCPU_CMD:
1013 			r = cb(ctx);
1014 			if (r)
1015 				return r;
1016 			break;
1017 		case mmUVD_ENGINE_CNTL:
1018 		case mmUVD_NO_OP:
1019 			break;
1020 		default:
1021 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1022 			return -EINVAL;
1023 		}
1024 		ctx->idx++;
1025 	}
1026 	return 0;
1027 }
1028 
1029 /**
1030  * amdgpu_uvd_cs_packets - parse UVD packets
1031  *
1032  * @ctx: UVD parser context
1033  * @cb: callback function
1034  *
1035  * Parse the command stream packets.
1036  */
amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))1037 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1038 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1039 {
1040 	int r;
1041 
1042 	for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1043 		uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1044 		unsigned type = CP_PACKET_GET_TYPE(cmd);
1045 		switch (type) {
1046 		case PACKET_TYPE0:
1047 			ctx->reg = CP_PACKET0_GET_REG(cmd);
1048 			ctx->count = CP_PACKET_GET_COUNT(cmd);
1049 			r = amdgpu_uvd_cs_reg(ctx, cb);
1050 			if (r)
1051 				return r;
1052 			break;
1053 		case PACKET_TYPE2:
1054 			++ctx->idx;
1055 			break;
1056 		default:
1057 			DRM_ERROR("Unknown packet type %d !\n", type);
1058 			return -EINVAL;
1059 		}
1060 	}
1061 	return 0;
1062 }
1063 
1064 /**
1065  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1066  *
1067  * @parser: Command submission parser context
1068  * @job: the job to parse
1069  * @ib: the IB to patch
1070  *
1071  * Parse the command stream, patch in addresses as necessary.
1072  */
amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser * parser,struct amdgpu_job * job,struct amdgpu_ib * ib)1073 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1074 			     struct amdgpu_job *job,
1075 			     struct amdgpu_ib *ib)
1076 {
1077 	struct amdgpu_uvd_cs_ctx ctx = {};
1078 	unsigned buf_sizes[] = {
1079 		[0x00000000]	=	2048,
1080 		[0x00000001]	=	0xFFFFFFFF,
1081 		[0x00000002]	=	0xFFFFFFFF,
1082 		[0x00000003]	=	2048,
1083 		[0x00000004]	=	0xFFFFFFFF,
1084 	};
1085 	int r;
1086 
1087 	job->vm = NULL;
1088 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
1089 
1090 	if (ib->length_dw % 16) {
1091 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1092 			  ib->length_dw);
1093 		return -EINVAL;
1094 	}
1095 
1096 	ctx.parser = parser;
1097 	ctx.buf_sizes = buf_sizes;
1098 	ctx.ib = ib;
1099 
1100 	/* first round only required on chips without UVD 64 bit address support */
1101 	if (!parser->adev->uvd.address_64_bit) {
1102 		/* first round, make sure the buffers are actually in the UVD segment */
1103 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1104 		if (r)
1105 			return r;
1106 	}
1107 
1108 	/* second round, patch buffer addresses into the command stream */
1109 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1110 	if (r)
1111 		return r;
1112 
1113 	if (!ctx.has_msg_cmd) {
1114 		DRM_ERROR("UVD-IBs need a msg command!\n");
1115 		return -EINVAL;
1116 	}
1117 
1118 	return 0;
1119 }
1120 
amdgpu_uvd_send_msg(struct amdgpu_ring * ring,struct amdgpu_bo * bo,bool direct,struct dma_fence ** fence)1121 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1122 			       bool direct, struct dma_fence **fence)
1123 {
1124 	struct amdgpu_device *adev = ring->adev;
1125 	struct dma_fence *f = NULL;
1126 	struct amdgpu_job *job;
1127 	struct amdgpu_ib *ib;
1128 	uint32_t data[4];
1129 	uint64_t addr;
1130 	long r;
1131 	int i;
1132 	unsigned offset_idx = 0;
1133 	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1134 
1135 	r = amdgpu_job_alloc_with_ib(adev, 64, direct ? AMDGPU_IB_POOL_DIRECT :
1136 				     AMDGPU_IB_POOL_DELAYED, &job);
1137 	if (r)
1138 		return r;
1139 
1140 	if (adev->asic_type >= CHIP_VEGA10) {
1141 		offset_idx = 1 + ring->me;
1142 		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1143 		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1144 	}
1145 
1146 	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1147 	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1148 	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1149 	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1150 
1151 	ib = &job->ibs[0];
1152 	addr = amdgpu_bo_gpu_offset(bo);
1153 	ib->ptr[0] = data[0];
1154 	ib->ptr[1] = addr;
1155 	ib->ptr[2] = data[1];
1156 	ib->ptr[3] = addr >> 32;
1157 	ib->ptr[4] = data[2];
1158 	ib->ptr[5] = 0;
1159 	for (i = 6; i < 16; i += 2) {
1160 		ib->ptr[i] = data[3];
1161 		ib->ptr[i+1] = 0;
1162 	}
1163 	ib->length_dw = 16;
1164 
1165 	if (direct) {
1166 		r = dma_resv_wait_timeout(bo->tbo.base.resv,
1167 					  DMA_RESV_USAGE_KERNEL, false,
1168 					  msecs_to_jiffies(10));
1169 		if (r == 0)
1170 			r = -ETIMEDOUT;
1171 		if (r < 0)
1172 			goto err_free;
1173 
1174 		r = amdgpu_job_submit_direct(job, ring, &f);
1175 		if (r)
1176 			goto err_free;
1177 	} else {
1178 		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1179 				     AMDGPU_SYNC_ALWAYS,
1180 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1181 		if (r)
1182 			goto err_free;
1183 
1184 		r = amdgpu_job_submit(job, &adev->uvd.entity,
1185 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1186 		if (r)
1187 			goto err_free;
1188 	}
1189 
1190 	amdgpu_bo_reserve(bo, true);
1191 	amdgpu_bo_fence(bo, f, false);
1192 	amdgpu_bo_unreserve(bo);
1193 
1194 	if (fence)
1195 		*fence = dma_fence_get(f);
1196 	dma_fence_put(f);
1197 
1198 	return 0;
1199 
1200 err_free:
1201 	amdgpu_job_free(job);
1202 	return r;
1203 }
1204 
1205 /* multiple fence commands without any stream commands in between can
1206    crash the vcpu so just try to emmit a dummy create/destroy msg to
1207    avoid this */
amdgpu_uvd_get_create_msg(struct amdgpu_ring * ring,uint32_t handle,struct dma_fence ** fence)1208 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1209 			      struct dma_fence **fence)
1210 {
1211 	struct amdgpu_device *adev = ring->adev;
1212 	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1213 	uint32_t *msg;
1214 	int i;
1215 
1216 	msg = amdgpu_bo_kptr(bo);
1217 	/* stitch together an UVD create msg */
1218 	msg[0] = cpu_to_le32(0x00000de4);
1219 	msg[1] = cpu_to_le32(0x00000000);
1220 	msg[2] = cpu_to_le32(handle);
1221 	msg[3] = cpu_to_le32(0x00000000);
1222 	msg[4] = cpu_to_le32(0x00000000);
1223 	msg[5] = cpu_to_le32(0x00000000);
1224 	msg[6] = cpu_to_le32(0x00000000);
1225 	msg[7] = cpu_to_le32(0x00000780);
1226 	msg[8] = cpu_to_le32(0x00000440);
1227 	msg[9] = cpu_to_le32(0x00000000);
1228 	msg[10] = cpu_to_le32(0x01b37000);
1229 	for (i = 11; i < 1024; ++i)
1230 		msg[i] = cpu_to_le32(0x0);
1231 
1232 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1233 
1234 }
1235 
amdgpu_uvd_get_destroy_msg(struct amdgpu_ring * ring,uint32_t handle,bool direct,struct dma_fence ** fence)1236 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1237 			       bool direct, struct dma_fence **fence)
1238 {
1239 	struct amdgpu_device *adev = ring->adev;
1240 	struct amdgpu_bo *bo = NULL;
1241 	uint32_t *msg;
1242 	int r, i;
1243 
1244 	if (direct) {
1245 		bo = adev->uvd.ib_bo;
1246 	} else {
1247 		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1248 		if (r)
1249 			return r;
1250 	}
1251 
1252 	msg = amdgpu_bo_kptr(bo);
1253 	/* stitch together an UVD destroy msg */
1254 	msg[0] = cpu_to_le32(0x00000de4);
1255 	msg[1] = cpu_to_le32(0x00000002);
1256 	msg[2] = cpu_to_le32(handle);
1257 	msg[3] = cpu_to_le32(0x00000000);
1258 	for (i = 4; i < 1024; ++i)
1259 		msg[i] = cpu_to_le32(0x0);
1260 
1261 	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1262 
1263 	if (!direct)
1264 		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1265 
1266 	return r;
1267 }
1268 
amdgpu_uvd_idle_work_handler(struct work_struct * work)1269 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1270 {
1271 	struct amdgpu_device *adev =
1272 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1273 	unsigned fences = 0, i, j;
1274 
1275 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1276 		if (adev->uvd.harvest_config & (1 << i))
1277 			continue;
1278 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1279 		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1280 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1281 		}
1282 	}
1283 
1284 	if (fences == 0) {
1285 		if (adev->pm.dpm_enabled) {
1286 			amdgpu_dpm_enable_uvd(adev, false);
1287 		} else {
1288 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1289 			/* shutdown the UVD block */
1290 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1291 							       AMD_PG_STATE_GATE);
1292 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1293 							       AMD_CG_STATE_GATE);
1294 		}
1295 	} else {
1296 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1297 	}
1298 }
1299 
amdgpu_uvd_ring_begin_use(struct amdgpu_ring * ring)1300 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1301 {
1302 	struct amdgpu_device *adev = ring->adev;
1303 	bool set_clocks;
1304 
1305 	if (amdgpu_sriov_vf(adev))
1306 		return;
1307 
1308 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1309 	if (set_clocks) {
1310 		if (adev->pm.dpm_enabled) {
1311 			amdgpu_dpm_enable_uvd(adev, true);
1312 		} else {
1313 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1314 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1315 							       AMD_CG_STATE_UNGATE);
1316 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1317 							       AMD_PG_STATE_UNGATE);
1318 		}
1319 	}
1320 }
1321 
amdgpu_uvd_ring_end_use(struct amdgpu_ring * ring)1322 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1323 {
1324 	if (!amdgpu_sriov_vf(ring->adev))
1325 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1326 }
1327 
1328 /**
1329  * amdgpu_uvd_ring_test_ib - test ib execution
1330  *
1331  * @ring: amdgpu_ring pointer
1332  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1333  *
1334  * Test if we can successfully execute an IB
1335  */
amdgpu_uvd_ring_test_ib(struct amdgpu_ring * ring,long timeout)1336 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1337 {
1338 	struct dma_fence *fence;
1339 	long r;
1340 
1341 	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1342 	if (r)
1343 		goto error;
1344 
1345 	r = dma_fence_wait_timeout(fence, false, timeout);
1346 	dma_fence_put(fence);
1347 	if (r == 0)
1348 		r = -ETIMEDOUT;
1349 	if (r < 0)
1350 		goto error;
1351 
1352 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1353 	if (r)
1354 		goto error;
1355 
1356 	r = dma_fence_wait_timeout(fence, false, timeout);
1357 	if (r == 0)
1358 		r = -ETIMEDOUT;
1359 	else if (r > 0)
1360 		r = 0;
1361 
1362 	dma_fence_put(fence);
1363 
1364 error:
1365 	return r;
1366 }
1367 
1368 /**
1369  * amdgpu_uvd_used_handles - returns used UVD handles
1370  *
1371  * @adev: amdgpu_device pointer
1372  *
1373  * Returns the number of UVD handles in use
1374  */
amdgpu_uvd_used_handles(struct amdgpu_device * adev)1375 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1376 {
1377 	unsigned i;
1378 	uint32_t used_handles = 0;
1379 
1380 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1381 		/*
1382 		 * Handles can be freed in any order, and not
1383 		 * necessarily linear. So we need to count
1384 		 * all non-zero handles.
1385 		 */
1386 		if (atomic_read(&adev->uvd.handles[i]))
1387 			used_handles++;
1388 	}
1389 
1390 	return used_handles;
1391 }
1392