xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision 8cbd01ba9c38eb16f3a572300da486ac544519b7)
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_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 int 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 = amdgpu_ucode_request(adev, &adev->uvd.fw, "%s", fw_name);
264 	if (r) {
265 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
266 			fw_name);
267 		amdgpu_ucode_release(&adev->uvd.fw);
268 		return r;
269 	}
270 
271 	/* Set the default UVD handles that the firmware can handle */
272 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
273 
274 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
275 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
276 
277 	if (adev->asic_type < CHIP_VEGA20) {
278 		unsigned int version_major, version_minor;
279 
280 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
281 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
282 		DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
283 			version_major, version_minor, family_id);
284 
285 		/*
286 		 * Limit the number of UVD handles depending on microcode major
287 		 * and minor versions. The firmware version which has 40 UVD
288 		 * instances support is 1.80. So all subsequent versions should
289 		 * also have the same support.
290 		 */
291 		if ((version_major > 0x01) ||
292 		    ((version_major == 0x01) && (version_minor >= 0x50)))
293 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
294 
295 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
296 					(family_id << 8));
297 
298 		if ((adev->asic_type == CHIP_POLARIS10 ||
299 		     adev->asic_type == CHIP_POLARIS11) &&
300 		    (adev->uvd.fw_version < FW_1_66_16))
301 			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
302 				  version_major, version_minor);
303 	} else {
304 		unsigned int enc_major, enc_minor, dec_minor;
305 
306 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
307 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
308 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
309 		DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
310 			enc_major, enc_minor, dec_minor, family_id);
311 
312 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
313 
314 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
315 	}
316 
317 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
318 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
319 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
320 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
321 
322 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
323 		if (adev->uvd.harvest_config & (1 << j))
324 			continue;
325 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
326 					    AMDGPU_GEM_DOMAIN_VRAM |
327 					    AMDGPU_GEM_DOMAIN_GTT,
328 					    &adev->uvd.inst[j].vcpu_bo,
329 					    &adev->uvd.inst[j].gpu_addr,
330 					    &adev->uvd.inst[j].cpu_addr);
331 		if (r) {
332 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
333 			return r;
334 		}
335 	}
336 
337 	for (i = 0; i < adev->uvd.max_handles; ++i) {
338 		atomic_set(&adev->uvd.handles[i], 0);
339 		adev->uvd.filp[i] = NULL;
340 	}
341 
342 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
343 	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
344 		adev->uvd.address_64_bit = true;
345 
346 	r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
347 	if (r)
348 		return r;
349 
350 	switch (adev->asic_type) {
351 	case CHIP_TONGA:
352 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
353 		break;
354 	case CHIP_CARRIZO:
355 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
356 		break;
357 	case CHIP_FIJI:
358 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
359 		break;
360 	case CHIP_STONEY:
361 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
362 		break;
363 	default:
364 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
365 	}
366 
367 	return 0;
368 }
369 
amdgpu_uvd_sw_fini(struct amdgpu_device * adev)370 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
371 {
372 	void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
373 	int i, j;
374 
375 	drm_sched_entity_destroy(&adev->uvd.entity);
376 
377 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
378 		if (adev->uvd.harvest_config & (1 << j))
379 			continue;
380 		kvfree(adev->uvd.inst[j].saved_bo);
381 
382 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
383 				      &adev->uvd.inst[j].gpu_addr,
384 				      (void **)&adev->uvd.inst[j].cpu_addr);
385 
386 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
387 
388 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
389 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
390 	}
391 	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
392 	amdgpu_ucode_release(&adev->uvd.fw);
393 
394 	return 0;
395 }
396 
397 /**
398  * amdgpu_uvd_entity_init - init entity
399  *
400  * @adev: amdgpu_device pointer
401  * @ring: amdgpu_ring pointer to check
402  *
403  * Initialize the entity used for handle management in the kernel driver.
404  */
amdgpu_uvd_entity_init(struct amdgpu_device * adev,struct amdgpu_ring * ring)405 int amdgpu_uvd_entity_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
406 {
407 	if (ring == &adev->uvd.inst[0].ring) {
408 		struct drm_gpu_scheduler *sched = &ring->sched;
409 		int r;
410 
411 		r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
412 					  &sched, 1, NULL);
413 		if (r) {
414 			DRM_ERROR("Failed setting up UVD kernel entity.\n");
415 			return r;
416 		}
417 	}
418 
419 	return 0;
420 }
421 
amdgpu_uvd_prepare_suspend(struct amdgpu_device * adev)422 int amdgpu_uvd_prepare_suspend(struct amdgpu_device *adev)
423 {
424 	unsigned int size;
425 	void *ptr;
426 	int i, j, idx;
427 
428 	cancel_delayed_work_sync(&adev->uvd.idle_work);
429 
430 	/* only valid for physical mode */
431 	if (adev->asic_type < CHIP_POLARIS10) {
432 		for (i = 0; i < adev->uvd.max_handles; ++i)
433 			if (atomic_read(&adev->uvd.handles[i]))
434 				break;
435 
436 		if (i == adev->uvd.max_handles)
437 			return 0;
438 	}
439 
440 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
441 		if (adev->uvd.harvest_config & (1 << j))
442 			continue;
443 		if (adev->uvd.inst[j].vcpu_bo == NULL)
444 			continue;
445 
446 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
447 		ptr = adev->uvd.inst[j].cpu_addr;
448 
449 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
450 		if (!adev->uvd.inst[j].saved_bo)
451 			return -ENOMEM;
452 
453 		if (drm_dev_enter(adev_to_drm(adev), &idx)) {
454 			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
455 			if (amdgpu_ras_intr_triggered())
456 				memset(adev->uvd.inst[j].saved_bo, 0, size);
457 			else
458 				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
459 
460 			drm_dev_exit(idx);
461 		}
462 	}
463 
464 	return 0;
465 }
466 
amdgpu_uvd_suspend(struct amdgpu_device * adev)467 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
468 {
469 	if (amdgpu_ras_intr_triggered())
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 int 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 int 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 
551 	for (i = 0; i < abo->placement.num_placement; ++i) {
552 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
553 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
554 		if (abo->placements[i].mem_type == TTM_PL_VRAM)
555 			abo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
556 	}
557 }
558 
amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx * ctx)559 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
560 {
561 	uint32_t lo, hi;
562 	uint64_t addr;
563 
564 	lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
565 	hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
566 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
567 
568 	return addr;
569 }
570 
571 /**
572  * amdgpu_uvd_cs_pass1 - first parsing round
573  *
574  * @ctx: UVD parser context
575  *
576  * Make sure UVD message and feedback buffers are in VRAM and
577  * nobody is violating an 256MB boundary.
578  */
amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx * ctx)579 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
580 {
581 	struct ttm_operation_ctx tctx = { false, false };
582 	struct amdgpu_bo_va_mapping *mapping;
583 	struct amdgpu_bo *bo;
584 	uint32_t cmd;
585 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
586 	int r = 0;
587 
588 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
589 	if (r) {
590 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
591 		return r;
592 	}
593 
594 	if (!ctx->parser->adev->uvd.address_64_bit) {
595 		/* check if it's a message or feedback command */
596 		cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
597 		if (cmd == 0x0 || cmd == 0x3) {
598 			/* yes, force it into VRAM */
599 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
600 
601 			amdgpu_bo_placement_from_domain(bo, domain);
602 		}
603 		amdgpu_uvd_force_into_uvd_segment(bo);
604 
605 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
606 	}
607 
608 	return r;
609 }
610 
611 /**
612  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
613  *
614  * @adev: amdgpu_device pointer
615  * @msg: pointer to message structure
616  * @buf_sizes: placeholder to put the different buffer lengths
617  *
618  * Peek into the decode message and calculate the necessary buffer sizes.
619  */
amdgpu_uvd_cs_msg_decode(struct amdgpu_device * adev,uint32_t * msg,unsigned int buf_sizes[])620 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
621 	unsigned int buf_sizes[])
622 {
623 	unsigned int stream_type = msg[4];
624 	unsigned int width = msg[6];
625 	unsigned int height = msg[7];
626 	unsigned int dpb_size = msg[9];
627 	unsigned int pitch = msg[28];
628 	unsigned int level = msg[57];
629 
630 	unsigned int width_in_mb = width / 16;
631 	unsigned int height_in_mb = ALIGN(height / 16, 2);
632 	unsigned int fs_in_mb = width_in_mb * height_in_mb;
633 
634 	unsigned int image_size, tmp, min_dpb_size, num_dpb_buffer;
635 	unsigned int min_ctx_size = ~0;
636 
637 	image_size = width * height;
638 	image_size += image_size / 2;
639 	image_size = ALIGN(image_size, 1024);
640 
641 	switch (stream_type) {
642 	case 0: /* H264 */
643 		switch (level) {
644 		case 30:
645 			num_dpb_buffer = 8100 / fs_in_mb;
646 			break;
647 		case 31:
648 			num_dpb_buffer = 18000 / fs_in_mb;
649 			break;
650 		case 32:
651 			num_dpb_buffer = 20480 / fs_in_mb;
652 			break;
653 		case 41:
654 			num_dpb_buffer = 32768 / fs_in_mb;
655 			break;
656 		case 42:
657 			num_dpb_buffer = 34816 / fs_in_mb;
658 			break;
659 		case 50:
660 			num_dpb_buffer = 110400 / fs_in_mb;
661 			break;
662 		case 51:
663 			num_dpb_buffer = 184320 / fs_in_mb;
664 			break;
665 		default:
666 			num_dpb_buffer = 184320 / fs_in_mb;
667 			break;
668 		}
669 		num_dpb_buffer++;
670 		if (num_dpb_buffer > 17)
671 			num_dpb_buffer = 17;
672 
673 		/* reference picture buffer */
674 		min_dpb_size = image_size * num_dpb_buffer;
675 
676 		/* macroblock context buffer */
677 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
678 
679 		/* IT surface buffer */
680 		min_dpb_size += width_in_mb * height_in_mb * 32;
681 		break;
682 
683 	case 1: /* VC1 */
684 
685 		/* reference picture buffer */
686 		min_dpb_size = image_size * 3;
687 
688 		/* CONTEXT_BUFFER */
689 		min_dpb_size += width_in_mb * height_in_mb * 128;
690 
691 		/* IT surface buffer */
692 		min_dpb_size += width_in_mb * 64;
693 
694 		/* DB surface buffer */
695 		min_dpb_size += width_in_mb * 128;
696 
697 		/* BP */
698 		tmp = max(width_in_mb, height_in_mb);
699 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
700 		break;
701 
702 	case 3: /* MPEG2 */
703 
704 		/* reference picture buffer */
705 		min_dpb_size = image_size * 3;
706 		break;
707 
708 	case 4: /* MPEG4 */
709 
710 		/* reference picture buffer */
711 		min_dpb_size = image_size * 3;
712 
713 		/* CM */
714 		min_dpb_size += width_in_mb * height_in_mb * 64;
715 
716 		/* IT surface buffer */
717 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
718 		break;
719 
720 	case 7: /* H264 Perf */
721 		switch (level) {
722 		case 30:
723 			num_dpb_buffer = 8100 / fs_in_mb;
724 			break;
725 		case 31:
726 			num_dpb_buffer = 18000 / fs_in_mb;
727 			break;
728 		case 32:
729 			num_dpb_buffer = 20480 / fs_in_mb;
730 			break;
731 		case 41:
732 			num_dpb_buffer = 32768 / fs_in_mb;
733 			break;
734 		case 42:
735 			num_dpb_buffer = 34816 / fs_in_mb;
736 			break;
737 		case 50:
738 			num_dpb_buffer = 110400 / fs_in_mb;
739 			break;
740 		case 51:
741 			num_dpb_buffer = 184320 / fs_in_mb;
742 			break;
743 		default:
744 			num_dpb_buffer = 184320 / fs_in_mb;
745 			break;
746 		}
747 		num_dpb_buffer++;
748 		if (num_dpb_buffer > 17)
749 			num_dpb_buffer = 17;
750 
751 		/* reference picture buffer */
752 		min_dpb_size = image_size * num_dpb_buffer;
753 
754 		if (!adev->uvd.use_ctx_buf) {
755 			/* macroblock context buffer */
756 			min_dpb_size +=
757 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
758 
759 			/* IT surface buffer */
760 			min_dpb_size += width_in_mb * height_in_mb * 32;
761 		} else {
762 			/* macroblock context buffer */
763 			min_ctx_size =
764 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
765 		}
766 		break;
767 
768 	case 8: /* MJPEG */
769 		min_dpb_size = 0;
770 		break;
771 
772 	case 16: /* H265 */
773 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
774 		image_size = ALIGN(image_size, 256);
775 
776 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
777 		min_dpb_size = image_size * num_dpb_buffer;
778 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
779 					   * 16 * num_dpb_buffer + 52 * 1024;
780 		break;
781 
782 	default:
783 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
784 		return -EINVAL;
785 	}
786 
787 	if (width > pitch) {
788 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
789 		return -EINVAL;
790 	}
791 
792 	if (dpb_size < min_dpb_size) {
793 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
794 			  dpb_size, min_dpb_size);
795 		return -EINVAL;
796 	}
797 
798 	buf_sizes[0x1] = dpb_size;
799 	buf_sizes[0x2] = image_size;
800 	buf_sizes[0x4] = min_ctx_size;
801 	/* store image width to adjust nb memory pstate */
802 	adev->uvd.decode_image_width = width;
803 	return 0;
804 }
805 
806 /**
807  * amdgpu_uvd_cs_msg - handle UVD message
808  *
809  * @ctx: UVD parser context
810  * @bo: buffer object containing the message
811  * @offset: offset into the buffer object
812  *
813  * Peek into the UVD message and extract the session id.
814  * Make sure that we don't open up to many sessions.
815  */
amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx * ctx,struct amdgpu_bo * bo,unsigned int offset)816 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
817 			     struct amdgpu_bo *bo, unsigned int offset)
818 {
819 	struct amdgpu_device *adev = ctx->parser->adev;
820 	int32_t *msg, msg_type, handle;
821 	void *ptr;
822 	long r;
823 	int i;
824 
825 	if (offset & 0x3F) {
826 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
827 		return -EINVAL;
828 	}
829 
830 	r = amdgpu_bo_kmap(bo, &ptr);
831 	if (r) {
832 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
833 		return r;
834 	}
835 
836 	msg = ptr + offset;
837 
838 	msg_type = msg[1];
839 	handle = msg[2];
840 
841 	if (handle == 0) {
842 		amdgpu_bo_kunmap(bo);
843 		DRM_ERROR("Invalid UVD handle!\n");
844 		return -EINVAL;
845 	}
846 
847 	switch (msg_type) {
848 	case 0:
849 		/* it's a create msg, calc image size (width * height) */
850 		amdgpu_bo_kunmap(bo);
851 
852 		/* try to alloc a new handle */
853 		for (i = 0; i < adev->uvd.max_handles; ++i) {
854 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
855 				DRM_ERROR(")Handle 0x%x already in use!\n",
856 					  handle);
857 				return -EINVAL;
858 			}
859 
860 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
861 				adev->uvd.filp[i] = ctx->parser->filp;
862 				return 0;
863 			}
864 		}
865 
866 		DRM_ERROR("No more free UVD handles!\n");
867 		return -ENOSPC;
868 
869 	case 1:
870 		/* it's a decode msg, calc buffer sizes */
871 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
872 		amdgpu_bo_kunmap(bo);
873 		if (r)
874 			return r;
875 
876 		/* validate the handle */
877 		for (i = 0; i < adev->uvd.max_handles; ++i) {
878 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
879 				if (adev->uvd.filp[i] != ctx->parser->filp) {
880 					DRM_ERROR("UVD handle collision detected!\n");
881 					return -EINVAL;
882 				}
883 				return 0;
884 			}
885 		}
886 
887 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
888 		return -ENOENT;
889 
890 	case 2:
891 		/* it's a destroy msg, free the handle */
892 		for (i = 0; i < adev->uvd.max_handles; ++i)
893 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
894 		amdgpu_bo_kunmap(bo);
895 		return 0;
896 
897 	default:
898 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
899 	}
900 
901 	amdgpu_bo_kunmap(bo);
902 	return -EINVAL;
903 }
904 
905 /**
906  * amdgpu_uvd_cs_pass2 - second parsing round
907  *
908  * @ctx: UVD parser context
909  *
910  * Patch buffer addresses, make sure buffer sizes are correct.
911  */
amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx * ctx)912 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
913 {
914 	struct amdgpu_bo_va_mapping *mapping;
915 	struct amdgpu_bo *bo;
916 	uint32_t cmd;
917 	uint64_t start, end;
918 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
919 	int r;
920 
921 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
922 	if (r) {
923 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
924 		return r;
925 	}
926 
927 	start = amdgpu_bo_gpu_offset(bo);
928 
929 	end = (mapping->last + 1 - mapping->start);
930 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
931 
932 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
933 	start += addr;
934 
935 	amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
936 	amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
937 
938 	cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
939 	if (cmd < 0x4) {
940 		if ((end - start) < ctx->buf_sizes[cmd]) {
941 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
942 				  (unsigned int)(end - start),
943 				  ctx->buf_sizes[cmd]);
944 			return -EINVAL;
945 		}
946 
947 	} else if (cmd == 0x206) {
948 		if ((end - start) < ctx->buf_sizes[4]) {
949 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
950 					  (unsigned int)(end - start),
951 					  ctx->buf_sizes[4]);
952 			return -EINVAL;
953 		}
954 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
955 		DRM_ERROR("invalid UVD command %X!\n", cmd);
956 		return -EINVAL;
957 	}
958 
959 	if (!ctx->parser->adev->uvd.address_64_bit) {
960 		if ((start >> 28) != ((end - 1) >> 28)) {
961 			DRM_ERROR("reloc %llx-%llx crossing 256MB boundary!\n",
962 				  start, end);
963 			return -EINVAL;
964 		}
965 
966 		if ((cmd == 0 || cmd == 0x3) &&
967 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
968 			DRM_ERROR("msg/fb buffer %llx-%llx out of 256MB segment!\n",
969 				  start, end);
970 			return -EINVAL;
971 		}
972 	}
973 
974 	if (cmd == 0) {
975 		ctx->has_msg_cmd = true;
976 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
977 		if (r)
978 			return r;
979 	} else if (!ctx->has_msg_cmd) {
980 		DRM_ERROR("Message needed before other commands are send!\n");
981 		return -EINVAL;
982 	}
983 
984 	return 0;
985 }
986 
987 /**
988  * amdgpu_uvd_cs_reg - parse register writes
989  *
990  * @ctx: UVD parser context
991  * @cb: callback function
992  *
993  * Parse the register writes, call cb on each complete command.
994  */
amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))995 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
996 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
997 {
998 	int i, r;
999 
1000 	ctx->idx++;
1001 	for (i = 0; i <= ctx->count; ++i) {
1002 		unsigned int reg = ctx->reg + i;
1003 
1004 		if (ctx->idx >= ctx->ib->length_dw) {
1005 			DRM_ERROR("Register command after end of CS!\n");
1006 			return -EINVAL;
1007 		}
1008 
1009 		switch (reg) {
1010 		case mmUVD_GPCOM_VCPU_DATA0:
1011 			ctx->data0 = ctx->idx;
1012 			break;
1013 		case mmUVD_GPCOM_VCPU_DATA1:
1014 			ctx->data1 = ctx->idx;
1015 			break;
1016 		case mmUVD_GPCOM_VCPU_CMD:
1017 			r = cb(ctx);
1018 			if (r)
1019 				return r;
1020 			break;
1021 		case mmUVD_ENGINE_CNTL:
1022 		case mmUVD_NO_OP:
1023 			break;
1024 		default:
1025 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1026 			return -EINVAL;
1027 		}
1028 		ctx->idx++;
1029 	}
1030 	return 0;
1031 }
1032 
1033 /**
1034  * amdgpu_uvd_cs_packets - parse UVD packets
1035  *
1036  * @ctx: UVD parser context
1037  * @cb: callback function
1038  *
1039  * Parse the command stream packets.
1040  */
amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))1041 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1042 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1043 {
1044 	int r;
1045 
1046 	for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1047 		uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1048 		unsigned int type = CP_PACKET_GET_TYPE(cmd);
1049 
1050 		switch (type) {
1051 		case PACKET_TYPE0:
1052 			ctx->reg = CP_PACKET0_GET_REG(cmd);
1053 			ctx->count = CP_PACKET_GET_COUNT(cmd);
1054 			r = amdgpu_uvd_cs_reg(ctx, cb);
1055 			if (r)
1056 				return r;
1057 			break;
1058 		case PACKET_TYPE2:
1059 			++ctx->idx;
1060 			break;
1061 		default:
1062 			DRM_ERROR("Unknown packet type %d !\n", type);
1063 			return -EINVAL;
1064 		}
1065 	}
1066 	return 0;
1067 }
1068 
1069 /**
1070  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1071  *
1072  * @parser: Command submission parser context
1073  * @job: the job to parse
1074  * @ib: the IB to patch
1075  *
1076  * Parse the command stream, patch in addresses as necessary.
1077  */
amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser * parser,struct amdgpu_job * job,struct amdgpu_ib * ib)1078 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1079 			     struct amdgpu_job *job,
1080 			     struct amdgpu_ib *ib)
1081 {
1082 	struct amdgpu_uvd_cs_ctx ctx = {};
1083 	unsigned int buf_sizes[] = {
1084 		[0x00000000]	=	2048,
1085 		[0x00000001]	=	0xFFFFFFFF,
1086 		[0x00000002]	=	0xFFFFFFFF,
1087 		[0x00000003]	=	2048,
1088 		[0x00000004]	=	0xFFFFFFFF,
1089 	};
1090 	int r;
1091 
1092 	job->vm = NULL;
1093 
1094 	if (ib->length_dw % 16) {
1095 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1096 			  ib->length_dw);
1097 		return -EINVAL;
1098 	}
1099 
1100 	ctx.parser = parser;
1101 	ctx.buf_sizes = buf_sizes;
1102 	ctx.ib = ib;
1103 
1104 	/* first round only required on chips without UVD 64 bit address support */
1105 	if (!parser->adev->uvd.address_64_bit) {
1106 		/* first round, make sure the buffers are actually in the UVD segment */
1107 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1108 		if (r)
1109 			return r;
1110 	}
1111 
1112 	/* second round, patch buffer addresses into the command stream */
1113 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1114 	if (r)
1115 		return r;
1116 
1117 	if (!ctx.has_msg_cmd) {
1118 		DRM_ERROR("UVD-IBs need a msg command!\n");
1119 		return -EINVAL;
1120 	}
1121 
1122 	return 0;
1123 }
1124 
amdgpu_uvd_send_msg(struct amdgpu_ring * ring,struct amdgpu_bo * bo,bool direct,struct dma_fence ** fence)1125 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1126 			       bool direct, struct dma_fence **fence)
1127 {
1128 	struct amdgpu_device *adev = ring->adev;
1129 	struct dma_fence *f = NULL;
1130 	uint32_t offset, data[4];
1131 	struct amdgpu_job *job;
1132 	struct amdgpu_ib *ib;
1133 	uint64_t addr;
1134 	int i, r;
1135 
1136 	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->uvd.entity,
1137 				     AMDGPU_FENCE_OWNER_UNDEFINED,
1138 				     64, direct ? AMDGPU_IB_POOL_DIRECT :
1139 				     AMDGPU_IB_POOL_DELAYED, &job);
1140 	if (r)
1141 		return r;
1142 
1143 	if (adev->asic_type >= CHIP_VEGA10)
1144 		offset = adev->reg_offset[UVD_HWIP][ring->me][1];
1145 	else
1146 		offset = UVD_BASE_SI;
1147 
1148 	data[0] = PACKET0(offset + UVD_GPCOM_VCPU_DATA0, 0);
1149 	data[1] = PACKET0(offset + UVD_GPCOM_VCPU_DATA1, 0);
1150 	data[2] = PACKET0(offset + UVD_GPCOM_VCPU_CMD, 0);
1151 	data[3] = PACKET0(offset + UVD_NO_OP, 0);
1152 
1153 	ib = &job->ibs[0];
1154 	addr = amdgpu_bo_gpu_offset(bo);
1155 	ib->ptr[0] = data[0];
1156 	ib->ptr[1] = addr;
1157 	ib->ptr[2] = data[1];
1158 	ib->ptr[3] = addr >> 32;
1159 	ib->ptr[4] = data[2];
1160 	ib->ptr[5] = 0;
1161 	for (i = 6; i < 16; i += 2) {
1162 		ib->ptr[i] = data[3];
1163 		ib->ptr[i+1] = 0;
1164 	}
1165 	ib->length_dw = 16;
1166 
1167 	if (direct) {
1168 		r = amdgpu_job_submit_direct(job, ring, &f);
1169 		if (r)
1170 			goto err_free;
1171 	} else {
1172 		r = drm_sched_job_add_resv_dependencies(&job->base,
1173 							bo->tbo.base.resv,
1174 							DMA_RESV_USAGE_KERNEL);
1175 		if (r)
1176 			goto err_free;
1177 
1178 		f = amdgpu_job_submit(job);
1179 	}
1180 
1181 	amdgpu_bo_reserve(bo, true);
1182 	amdgpu_bo_fence(bo, f, false);
1183 	amdgpu_bo_unreserve(bo);
1184 
1185 	if (fence)
1186 		*fence = dma_fence_get(f);
1187 	dma_fence_put(f);
1188 
1189 	return 0;
1190 
1191 err_free:
1192 	amdgpu_job_free(job);
1193 	return r;
1194 }
1195 
1196 /* multiple fence commands without any stream commands in between can
1197  * crash the vcpu so just try to emmit a dummy create/destroy msg to
1198  * avoid this
1199  */
amdgpu_uvd_get_create_msg(struct amdgpu_ring * ring,uint32_t handle,struct dma_fence ** fence)1200 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1201 			      struct dma_fence **fence)
1202 {
1203 	struct amdgpu_device *adev = ring->adev;
1204 	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1205 	uint32_t *msg;
1206 	int i;
1207 
1208 	msg = amdgpu_bo_kptr(bo);
1209 	/* stitch together an UVD create msg */
1210 	msg[0] = cpu_to_le32(0x00000de4);
1211 	msg[1] = cpu_to_le32(0x00000000);
1212 	msg[2] = cpu_to_le32(handle);
1213 	msg[3] = cpu_to_le32(0x00000000);
1214 	msg[4] = cpu_to_le32(0x00000000);
1215 	msg[5] = cpu_to_le32(0x00000000);
1216 	msg[6] = cpu_to_le32(0x00000000);
1217 	msg[7] = cpu_to_le32(0x00000780);
1218 	msg[8] = cpu_to_le32(0x00000440);
1219 	msg[9] = cpu_to_le32(0x00000000);
1220 	msg[10] = cpu_to_le32(0x01b37000);
1221 	for (i = 11; i < 1024; ++i)
1222 		msg[i] = cpu_to_le32(0x0);
1223 
1224 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1225 
1226 }
1227 
amdgpu_uvd_get_destroy_msg(struct amdgpu_ring * ring,uint32_t handle,bool direct,struct dma_fence ** fence)1228 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1229 			       bool direct, struct dma_fence **fence)
1230 {
1231 	struct amdgpu_device *adev = ring->adev;
1232 	struct amdgpu_bo *bo = NULL;
1233 	uint32_t *msg;
1234 	int r, i;
1235 
1236 	if (direct) {
1237 		bo = adev->uvd.ib_bo;
1238 	} else {
1239 		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1240 		if (r)
1241 			return r;
1242 	}
1243 
1244 	msg = amdgpu_bo_kptr(bo);
1245 	/* stitch together an UVD destroy msg */
1246 	msg[0] = cpu_to_le32(0x00000de4);
1247 	msg[1] = cpu_to_le32(0x00000002);
1248 	msg[2] = cpu_to_le32(handle);
1249 	msg[3] = cpu_to_le32(0x00000000);
1250 	for (i = 4; i < 1024; ++i)
1251 		msg[i] = cpu_to_le32(0x0);
1252 
1253 	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1254 
1255 	if (!direct)
1256 		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1257 
1258 	return r;
1259 }
1260 
amdgpu_uvd_idle_work_handler(struct work_struct * work)1261 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1262 {
1263 	struct amdgpu_device *adev =
1264 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1265 	unsigned int fences = 0, i, j;
1266 
1267 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1268 		if (adev->uvd.harvest_config & (1 << i))
1269 			continue;
1270 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1271 		for (j = 0; j < adev->uvd.num_enc_rings; ++j)
1272 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1273 	}
1274 
1275 	if (fences == 0) {
1276 		if (adev->pm.dpm_enabled) {
1277 			amdgpu_dpm_enable_uvd(adev, false);
1278 		} else {
1279 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1280 			/* shutdown the UVD block */
1281 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1282 							       AMD_PG_STATE_GATE);
1283 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1284 							       AMD_CG_STATE_GATE);
1285 		}
1286 	} else {
1287 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1288 	}
1289 }
1290 
amdgpu_uvd_ring_begin_use(struct amdgpu_ring * ring)1291 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1292 {
1293 	struct amdgpu_device *adev = ring->adev;
1294 	bool set_clocks;
1295 
1296 	if (amdgpu_sriov_vf(adev))
1297 		return;
1298 
1299 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1300 	if (set_clocks) {
1301 		if (adev->pm.dpm_enabled) {
1302 			amdgpu_dpm_enable_uvd(adev, true);
1303 		} else {
1304 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1305 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1306 							       AMD_CG_STATE_UNGATE);
1307 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1308 							       AMD_PG_STATE_UNGATE);
1309 		}
1310 	}
1311 }
1312 
amdgpu_uvd_ring_end_use(struct amdgpu_ring * ring)1313 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1314 {
1315 	if (!amdgpu_sriov_vf(ring->adev))
1316 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1317 }
1318 
1319 /**
1320  * amdgpu_uvd_ring_test_ib - test ib execution
1321  *
1322  * @ring: amdgpu_ring pointer
1323  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1324  *
1325  * Test if we can successfully execute an IB
1326  */
amdgpu_uvd_ring_test_ib(struct amdgpu_ring * ring,long timeout)1327 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1328 {
1329 	struct dma_fence *fence;
1330 	long r;
1331 
1332 	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1333 	if (r)
1334 		goto error;
1335 
1336 	r = dma_fence_wait_timeout(fence, false, timeout);
1337 	dma_fence_put(fence);
1338 	if (r == 0)
1339 		r = -ETIMEDOUT;
1340 	if (r < 0)
1341 		goto error;
1342 
1343 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1344 	if (r)
1345 		goto error;
1346 
1347 	r = dma_fence_wait_timeout(fence, false, timeout);
1348 	if (r == 0)
1349 		r = -ETIMEDOUT;
1350 	else if (r > 0)
1351 		r = 0;
1352 
1353 	dma_fence_put(fence);
1354 
1355 error:
1356 	return r;
1357 }
1358 
1359 /**
1360  * amdgpu_uvd_used_handles - returns used UVD handles
1361  *
1362  * @adev: amdgpu_device pointer
1363  *
1364  * Returns the number of UVD handles in use
1365  */
amdgpu_uvd_used_handles(struct amdgpu_device * adev)1366 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1367 {
1368 	unsigned int i;
1369 	uint32_t used_handles = 0;
1370 
1371 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1372 		/*
1373 		 * Handles can be freed in any order, and not
1374 		 * necessarily linear. So we need to count
1375 		 * all non-zero handles.
1376 		 */
1377 		if (atomic_read(&adev->uvd.handles[i]))
1378 			used_handles++;
1379 	}
1380 
1381 	return used_handles;
1382 }
1383