xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c (revision fab183d632628381b466a41479489541ac0e29a0)
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 
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_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 
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 	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 
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 	amdgpu_ucode_release(&adev->uvd.fw);
398 
399 	return 0;
400 }
401 
402 /**
403  * amdgpu_uvd_entity_init - init entity
404  *
405  * @adev: amdgpu_device pointer
406  * @ring: amdgpu_ring pointer to check
407  *
408  * Initialize the entity used for handle management in the kernel driver.
409  */
amdgpu_uvd_entity_init(struct amdgpu_device * adev,struct amdgpu_ring * ring)410 int amdgpu_uvd_entity_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
411 {
412 	if (ring == &adev->uvd.inst[0].ring) {
413 		struct drm_gpu_scheduler *sched = &ring->sched;
414 		int r;
415 
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 
424 	return 0;
425 }
426 
amdgpu_uvd_prepare_suspend(struct amdgpu_device * adev)427 int amdgpu_uvd_prepare_suspend(struct amdgpu_device *adev)
428 {
429 	unsigned int size;
430 	void *ptr;
431 	int i, j, idx;
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 (amdgpu_ras_intr_triggered())
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 	return 0;
470 }
471 
amdgpu_uvd_suspend(struct amdgpu_device * adev)472 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
473 {
474 	if (amdgpu_ras_intr_triggered())
475 		drm_warn(adev_to_drm(adev),
476 			"UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
477 
478 	return 0;
479 }
480 
amdgpu_uvd_resume(struct amdgpu_device * adev)481 int amdgpu_uvd_resume(struct amdgpu_device *adev)
482 {
483 	unsigned int size;
484 	void *ptr;
485 	int i, idx;
486 
487 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
488 		if (adev->uvd.harvest_config & (1 << i))
489 			continue;
490 		if (adev->uvd.inst[i].vcpu_bo == NULL)
491 			return -EINVAL;
492 
493 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
494 		ptr = adev->uvd.inst[i].cpu_addr;
495 
496 		if (adev->uvd.inst[i].saved_bo != NULL) {
497 			if (drm_dev_enter(adev_to_drm(adev), &idx)) {
498 				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
499 				drm_dev_exit(idx);
500 			}
501 			kvfree(adev->uvd.inst[i].saved_bo);
502 			adev->uvd.inst[i].saved_bo = NULL;
503 		} else {
504 			const struct common_firmware_header *hdr;
505 			unsigned int offset;
506 
507 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
508 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
509 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
510 				if (drm_dev_enter(adev_to_drm(adev), &idx)) {
511 					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
512 						    le32_to_cpu(hdr->ucode_size_bytes));
513 					drm_dev_exit(idx);
514 				}
515 				size -= le32_to_cpu(hdr->ucode_size_bytes);
516 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
517 			}
518 			memset_io(ptr, 0, size);
519 			/* to restore uvd fence seq */
520 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring, NULL);
521 		}
522 	}
523 	return 0;
524 }
525 
amdgpu_uvd_free_handles(struct amdgpu_device * adev,struct drm_file * filp)526 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
527 {
528 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
529 	int i, r;
530 
531 	for (i = 0; i < adev->uvd.max_handles; ++i) {
532 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
533 
534 		if (handle != 0 && adev->uvd.filp[i] == filp) {
535 			struct dma_fence *fence;
536 
537 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
538 						       &fence);
539 			if (r) {
540 				DRM_ERROR("Error destroying UVD %d!\n", r);
541 				continue;
542 			}
543 
544 			dma_fence_wait(fence, false);
545 			dma_fence_put(fence);
546 
547 			adev->uvd.filp[i] = NULL;
548 			atomic_set(&adev->uvd.handles[i], 0);
549 		}
550 	}
551 }
552 
amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo * bo)553 static void amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo *bo)
554 {
555 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
556 	struct amdgpu_bo *vcpu_bo = adev->uvd.inst[0].vcpu_bo;
557 	struct amdgpu_res_cursor vcpu_cur;
558 
559 	amdgpu_res_first(vcpu_bo->tbo.resource, 0,
560 			 amdgpu_bo_size(vcpu_bo), &vcpu_cur);
561 
562 	bo->placement.num_placement = 1;
563 	bo->placement.placement = &bo->placements[0];
564 	bo->placements[0].fpfn = ALIGN_DOWN(vcpu_cur.start, SZ_256M) >> PAGE_SHIFT;
565 	bo->placements[0].lpfn = bo->placements[0].fpfn + (SZ_256M >> PAGE_SHIFT);
566 	bo->placements[0].mem_type = vcpu_bo->tbo.resource->mem_type;
567 	if (bo->placements[0].mem_type == TTM_PL_VRAM)
568 		bo->placements[0].flags |= TTM_PL_FLAG_CONTIGUOUS;
569 }
570 
amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo * abo)571 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
572 {
573 	int i;
574 
575 	for (i = 0; i < abo->placement.num_placement; ++i) {
576 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
577 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
578 		if (abo->placements[i].mem_type == TTM_PL_VRAM)
579 			abo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
580 	}
581 }
582 
amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx * ctx)583 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
584 {
585 	uint32_t lo, hi;
586 	uint64_t addr;
587 
588 	lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
589 	hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
590 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
591 
592 	return addr;
593 }
594 
595 /**
596  * amdgpu_uvd_cs_pass1 - first parsing round
597  *
598  * @ctx: UVD parser context
599  *
600  * Make sure UVD message and feedback buffers are in VRAM and
601  * nobody is violating an 256MB boundary.
602  */
amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx * ctx)603 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
604 {
605 	struct ttm_operation_ctx tctx = { false, false };
606 	struct amdgpu_bo_va_mapping *mapping;
607 	struct amdgpu_bo *bo;
608 	uint32_t cmd;
609 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
610 	int r = 0;
611 
612 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
613 	if (r) {
614 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
615 		return r;
616 	}
617 
618 	if (!ctx->parser->adev->uvd.address_64_bit) {
619 		/* check if it's a message or feedback command */
620 		cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
621 		if (cmd == 0x0 || cmd == 0x3)
622 			amdgpu_uvd_force_into_vcpu_segment(bo);
623 		else
624 			amdgpu_uvd_force_into_uvd_segment(bo);
625 
626 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
627 	}
628 
629 	return r;
630 }
631 
632 /**
633  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
634  *
635  * @adev: amdgpu_device pointer
636  * @msg: pointer to message structure
637  * @buf_sizes: placeholder to put the different buffer lengths
638  *
639  * Peek into the decode message and calculate the necessary buffer sizes.
640  */
amdgpu_uvd_cs_msg_decode(struct amdgpu_device * adev,uint32_t * msg,unsigned int buf_sizes[])641 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
642 	unsigned int buf_sizes[])
643 {
644 	unsigned int stream_type = msg[4];
645 	unsigned int width = msg[6];
646 	unsigned int height = msg[7];
647 	unsigned int dpb_size = msg[9];
648 	unsigned int pitch = msg[28];
649 
650 	unsigned int width_in_mb = width / 16;
651 	unsigned int height_in_mb = ALIGN(height / 16, 2);
652 
653 	unsigned int image_size, tmp, min_dpb_size, num_dpb_buffer;
654 	unsigned int min_ctx_size = ~0;
655 
656 	/* Reject invalid dimensions */
657 	if (width < 16 || height < 16 || width > 4096 || height > 4096) {
658 		dev_WARN_ONCE(adev->dev, 1,
659 			      "Invalid UVD decoding dimensions (%dx%d)!\n",
660 			      width, height);
661 		return -EINVAL;
662 	}
663 
664 	image_size = width * height;
665 	image_size += image_size / 2;
666 	image_size = ALIGN(image_size, 1024);
667 
668 	switch (stream_type) {
669 	case 0: /* H264 */
670 		num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
671 		if (num_dpb_buffer > 17)
672 			return -EINVAL;
673 
674 		/* reference picture buffer */
675 		min_dpb_size = image_size * num_dpb_buffer;
676 
677 		/* macroblock context buffer */
678 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
679 
680 		/* IT surface buffer */
681 		min_dpb_size += width_in_mb * height_in_mb * 32;
682 		break;
683 
684 	case 1: /* VC1 */
685 
686 		/* reference picture buffer */
687 		min_dpb_size = image_size * 3;
688 
689 		/* CONTEXT_BUFFER */
690 		min_dpb_size += width_in_mb * height_in_mb * 128;
691 
692 		/* IT surface buffer */
693 		min_dpb_size += width_in_mb * 64;
694 
695 		/* DB surface buffer */
696 		min_dpb_size += width_in_mb * 128;
697 
698 		/* BP */
699 		tmp = max(width_in_mb, height_in_mb);
700 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
701 		break;
702 
703 	case 3: /* MPEG2 */
704 
705 		/* reference picture buffer */
706 		min_dpb_size = image_size * 3;
707 		break;
708 
709 	case 4: /* MPEG4 */
710 
711 		/* reference picture buffer */
712 		min_dpb_size = image_size * 3;
713 
714 		/* CM */
715 		min_dpb_size += width_in_mb * height_in_mb * 64;
716 
717 		/* IT surface buffer */
718 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
719 		break;
720 
721 	case 7: /* H264 Perf */
722 		num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
723 		if (num_dpb_buffer > 17)
724 			return -EINVAL;
725 
726 		/* reference picture buffer */
727 		min_dpb_size = image_size * num_dpb_buffer;
728 
729 		if (!adev->uvd.use_ctx_buf) {
730 			/* macroblock context buffer */
731 			min_dpb_size +=
732 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
733 
734 			/* IT surface buffer */
735 			min_dpb_size += width_in_mb * height_in_mb * 32;
736 		} else {
737 			/* macroblock context buffer */
738 			min_ctx_size =
739 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
740 		}
741 		break;
742 
743 	case 8: /* MJPEG */
744 		min_dpb_size = 0;
745 		break;
746 
747 	case 16: /* H265 */
748 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
749 		image_size = ALIGN(image_size, 256);
750 
751 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
752 		if (num_dpb_buffer > 17)
753 			return -EINVAL;
754 
755 		min_dpb_size = image_size * num_dpb_buffer;
756 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
757 					   * 16 * num_dpb_buffer + 52 * 1024;
758 		break;
759 
760 	default:
761 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
762 		return -EINVAL;
763 	}
764 
765 	if (width > pitch || pitch > 4096) {
766 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
767 		return -EINVAL;
768 	}
769 
770 	if (dpb_size < min_dpb_size) {
771 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
772 			  dpb_size, min_dpb_size);
773 		return -EINVAL;
774 	}
775 
776 	buf_sizes[0x1] = dpb_size;
777 	buf_sizes[0x2] = (pitch * height) * 3 / 2;
778 	buf_sizes[0x4] = min_ctx_size;
779 	/* store image width to adjust nb memory pstate */
780 	adev->uvd.decode_image_width = width;
781 	return 0;
782 }
783 
784 /**
785  * amdgpu_uvd_cs_msg - handle UVD message
786  *
787  * @ctx: UVD parser context
788  * @bo: buffer object containing the message
789  * @offset: offset into the buffer object
790  *
791  * Peek into the UVD message and extract the session id.
792  * Make sure that we don't open up to many sessions.
793  */
amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx * ctx,struct amdgpu_bo * bo,unsigned int offset)794 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
795 			     struct amdgpu_bo *bo, unsigned int offset)
796 {
797 	struct amdgpu_device *adev = ctx->parser->adev;
798 	int32_t *msg, msg_type, handle;
799 	void *ptr;
800 	long r;
801 	int i;
802 
803 	if (offset & 0x3F) {
804 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
805 		return -EINVAL;
806 	}
807 
808 	r = amdgpu_bo_kmap(bo, &ptr);
809 	if (r) {
810 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
811 		return r;
812 	}
813 
814 	msg = ptr + offset;
815 
816 	msg_type = msg[1];
817 	handle = msg[2];
818 
819 	if (handle == 0) {
820 		amdgpu_bo_kunmap(bo);
821 		DRM_ERROR("Invalid UVD handle!\n");
822 		return -EINVAL;
823 	}
824 
825 	switch (msg_type) {
826 	case 0:
827 		/* it's a create msg, calc image size (width * height) */
828 		amdgpu_bo_kunmap(bo);
829 
830 		/* try to alloc a new handle */
831 		for (i = 0; i < adev->uvd.max_handles; ++i) {
832 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
833 				DRM_ERROR(")Handle 0x%x already in use!\n",
834 					  handle);
835 				return -EINVAL;
836 			}
837 
838 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
839 				adev->uvd.filp[i] = ctx->parser->filp;
840 				return 0;
841 			}
842 		}
843 
844 		DRM_ERROR("No more free UVD handles!\n");
845 		return -ENOSPC;
846 
847 	case 1:
848 		/* it's a decode msg, calc buffer sizes */
849 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
850 		amdgpu_bo_kunmap(bo);
851 		if (r)
852 			return r;
853 
854 		/* validate the handle */
855 		for (i = 0; i < adev->uvd.max_handles; ++i) {
856 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
857 				if (adev->uvd.filp[i] != ctx->parser->filp) {
858 					DRM_ERROR("UVD handle collision detected!\n");
859 					return -EINVAL;
860 				}
861 				return 0;
862 			}
863 		}
864 
865 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
866 		return -ENOENT;
867 
868 	case 2:
869 		/* it's a destroy msg, free the handle */
870 		for (i = 0; i < adev->uvd.max_handles; ++i)
871 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
872 		amdgpu_bo_kunmap(bo);
873 		return 0;
874 
875 	default:
876 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
877 	}
878 
879 	amdgpu_bo_kunmap(bo);
880 	return -EINVAL;
881 }
882 
883 /**
884  * amdgpu_uvd_cs_pass2 - second parsing round
885  *
886  * @ctx: UVD parser context
887  *
888  * Patch buffer addresses, make sure buffer sizes are correct.
889  */
amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx * ctx)890 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
891 {
892 	struct amdgpu_bo_va_mapping *mapping;
893 	struct amdgpu_bo *bo;
894 	uint32_t cmd;
895 	uint64_t start, end;
896 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
897 	int r;
898 
899 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
900 	if (r) {
901 		DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
902 		return r;
903 	}
904 
905 	start = amdgpu_bo_gpu_offset(bo);
906 
907 	end = (mapping->last + 1 - mapping->start);
908 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
909 
910 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
911 	start += addr;
912 
913 	amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
914 	amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
915 
916 	cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
917 	if (cmd < 0x4) {
918 		if ((end - start) < ctx->buf_sizes[cmd]) {
919 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
920 				  (unsigned int)(end - start),
921 				  ctx->buf_sizes[cmd]);
922 			return -EINVAL;
923 		}
924 	} else if (cmd == 0x204 || cmd == 0x206) {
925 		unsigned int min_size = ctx->buf_sizes[cmd == 0x204 ? 5 : 4];
926 
927 		if ((end - start) < min_size) {
928 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
929 					  (unsigned int)(end - start),
930 					  min_size);
931 			return -EINVAL;
932 		}
933 	} else if ((cmd != 0x100)) {
934 		DRM_ERROR("invalid UVD command %X!\n", cmd);
935 		return -EINVAL;
936 	}
937 
938 	if (!ctx->parser->adev->uvd.address_64_bit) {
939 		if ((start >> 28) != ((end - 1) >> 28)) {
940 			DRM_ERROR("reloc %llx-%llx crossing 256MB boundary!\n",
941 				  start, end);
942 			return -EINVAL;
943 		}
944 
945 		if ((cmd == 0 || cmd == 0x3) &&
946 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
947 			DRM_ERROR("msg/fb buffer %llx-%llx out of 256MB segment!\n",
948 				  start, end);
949 			return -EINVAL;
950 		}
951 	}
952 
953 	if (cmd == 0) {
954 		ctx->has_msg_cmd = true;
955 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
956 		if (r)
957 			return r;
958 	} else if (!ctx->has_msg_cmd) {
959 		DRM_ERROR("Message needed before other commands are send!\n");
960 		return -EINVAL;
961 	}
962 
963 	return 0;
964 }
965 
966 /**
967  * amdgpu_uvd_cs_reg - parse register writes
968  *
969  * @ctx: UVD parser context
970  * @cb: callback function
971  *
972  * Parse the register writes, call cb on each complete command.
973  */
amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))974 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
975 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
976 {
977 	int i, r;
978 
979 	ctx->idx++;
980 	for (i = 0; i <= ctx->count; ++i) {
981 		unsigned int reg = ctx->reg + i;
982 
983 		if (ctx->idx >= ctx->ib->length_dw) {
984 			DRM_ERROR("Register command after end of CS!\n");
985 			return -EINVAL;
986 		}
987 
988 		switch (reg) {
989 		case mmUVD_GPCOM_VCPU_DATA0:
990 			ctx->data0 = ctx->idx;
991 			break;
992 		case mmUVD_GPCOM_VCPU_DATA1:
993 			ctx->data1 = ctx->idx;
994 			break;
995 		case mmUVD_GPCOM_VCPU_CMD:
996 			r = cb(ctx);
997 			if (r)
998 				return r;
999 			break;
1000 		case mmUVD_ENGINE_CNTL:
1001 		case mmUVD_NO_OP:
1002 			break;
1003 		default:
1004 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1005 			return -EINVAL;
1006 		}
1007 		ctx->idx++;
1008 	}
1009 	return 0;
1010 }
1011 
1012 /**
1013  * amdgpu_uvd_cs_packets - parse UVD packets
1014  *
1015  * @ctx: UVD parser context
1016  * @cb: callback function
1017  *
1018  * Parse the command stream packets.
1019  */
amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))1020 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1021 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1022 {
1023 	int r;
1024 
1025 	for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1026 		uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1027 		unsigned int type = CP_PACKET_GET_TYPE(cmd);
1028 
1029 		switch (type) {
1030 		case PACKET_TYPE0:
1031 			ctx->reg = CP_PACKET0_GET_REG(cmd);
1032 			ctx->count = CP_PACKET_GET_COUNT(cmd);
1033 			r = amdgpu_uvd_cs_reg(ctx, cb);
1034 			if (r)
1035 				return r;
1036 			break;
1037 		case PACKET_TYPE2:
1038 			++ctx->idx;
1039 			break;
1040 		default:
1041 			DRM_ERROR("Unknown packet type %d !\n", type);
1042 			return -EINVAL;
1043 		}
1044 	}
1045 	return 0;
1046 }
1047 
1048 /**
1049  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1050  *
1051  * @parser: Command submission parser context
1052  * @job: the job to parse
1053  * @ib: the IB to patch
1054  *
1055  * Parse the command stream, patch in addresses as necessary.
1056  */
amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser * parser,struct amdgpu_job * job,struct amdgpu_ib * ib)1057 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1058 			     struct amdgpu_job *job,
1059 			     struct amdgpu_ib *ib)
1060 {
1061 	struct amdgpu_uvd_cs_ctx ctx = {};
1062 	unsigned int buf_sizes[] = {
1063 		[0x00000000]	=	3556,
1064 		[0x00000001]	=	0xFFFFFFFF,
1065 		[0x00000002]	=	0xFFFFFFFF,
1066 		[0x00000003]	=	2048,
1067 		[0x00000004]	=	0xFFFFFFFF,
1068 		[0x00000005]	=	992,
1069 	};
1070 	int r;
1071 
1072 	job->vm = NULL;
1073 
1074 	if (ib->length_dw % 16) {
1075 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1076 			  ib->length_dw);
1077 		return -EINVAL;
1078 	}
1079 
1080 	ctx.parser = parser;
1081 	ctx.buf_sizes = buf_sizes;
1082 	ctx.ib = ib;
1083 
1084 	/* first round only required on chips without UVD 64 bit address support */
1085 	if (!parser->adev->uvd.address_64_bit) {
1086 		/* first round, make sure the buffers are actually in the UVD segment */
1087 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1088 		if (r)
1089 			return r;
1090 	}
1091 
1092 	/* second round, patch buffer addresses into the command stream */
1093 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1094 	if (r)
1095 		return r;
1096 
1097 	if (!ctx.has_msg_cmd) {
1098 		DRM_ERROR("UVD-IBs need a msg command!\n");
1099 		return -EINVAL;
1100 	}
1101 
1102 	return 0;
1103 }
1104 
amdgpu_uvd_send_msg(struct amdgpu_ring * ring,struct amdgpu_bo * bo,bool direct,struct dma_fence ** fence)1105 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1106 			       bool direct, struct dma_fence **fence)
1107 {
1108 	struct amdgpu_device *adev = ring->adev;
1109 	struct dma_fence *f = NULL;
1110 	uint32_t offset, data[4];
1111 	struct amdgpu_job *job;
1112 	struct amdgpu_ib *ib;
1113 	uint64_t addr;
1114 	int i, r;
1115 
1116 	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->uvd.entity,
1117 				     AMDGPU_FENCE_OWNER_UNDEFINED,
1118 				     64, direct ? AMDGPU_IB_POOL_DIRECT :
1119 				     AMDGPU_IB_POOL_DELAYED, &job,
1120 				     AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST);
1121 	if (r)
1122 		return r;
1123 
1124 	if (adev->asic_type >= CHIP_VEGA10)
1125 		offset = adev->reg_offset[UVD_HWIP][ring->me][1];
1126 	else
1127 		offset = UVD_BASE_SI;
1128 
1129 	data[0] = PACKET0(offset + UVD_GPCOM_VCPU_DATA0, 0);
1130 	data[1] = PACKET0(offset + UVD_GPCOM_VCPU_DATA1, 0);
1131 	data[2] = PACKET0(offset + UVD_GPCOM_VCPU_CMD, 0);
1132 	data[3] = PACKET0(offset + UVD_NO_OP, 0);
1133 
1134 	ib = &job->ibs[0];
1135 	addr = amdgpu_bo_gpu_offset(bo);
1136 	ib->ptr[0] = data[0];
1137 	ib->ptr[1] = addr;
1138 	ib->ptr[2] = data[1];
1139 	ib->ptr[3] = addr >> 32;
1140 	ib->ptr[4] = data[2];
1141 	ib->ptr[5] = 0;
1142 	for (i = 6; i < 16; i += 2) {
1143 		ib->ptr[i] = data[3];
1144 		ib->ptr[i+1] = 0;
1145 	}
1146 	ib->length_dw = 16;
1147 
1148 	if (direct) {
1149 		r = amdgpu_job_submit_direct(job, ring, &f);
1150 		if (r)
1151 			goto err_free;
1152 	} else {
1153 		r = drm_sched_job_add_resv_dependencies(&job->base,
1154 							bo->tbo.base.resv,
1155 							DMA_RESV_USAGE_KERNEL);
1156 		if (r)
1157 			goto err_free;
1158 
1159 		f = amdgpu_job_submit(job);
1160 	}
1161 
1162 	amdgpu_bo_reserve(bo, true);
1163 	amdgpu_bo_fence(bo, f, false);
1164 	amdgpu_bo_unreserve(bo);
1165 
1166 	if (fence)
1167 		*fence = dma_fence_get(f);
1168 	dma_fence_put(f);
1169 
1170 	return 0;
1171 
1172 err_free:
1173 	amdgpu_job_free(job);
1174 	return r;
1175 }
1176 
1177 /* multiple fence commands without any stream commands in between can
1178  * crash the vcpu so just try to emmit a dummy create/destroy msg to
1179  * avoid this
1180  */
amdgpu_uvd_get_create_msg(struct amdgpu_ring * ring,uint32_t handle,struct dma_fence ** fence)1181 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1182 			      struct dma_fence **fence)
1183 {
1184 	struct amdgpu_device *adev = ring->adev;
1185 	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1186 	uint32_t *msg;
1187 	int i;
1188 
1189 	msg = amdgpu_bo_kptr(bo);
1190 	/* stitch together an UVD create msg */
1191 	msg[0] = cpu_to_le32(0x00000de4);
1192 	msg[1] = cpu_to_le32(0x00000000);
1193 	msg[2] = cpu_to_le32(handle);
1194 	msg[3] = cpu_to_le32(0x00000000);
1195 	msg[4] = cpu_to_le32(0x00000000);
1196 	msg[5] = cpu_to_le32(0x00000000);
1197 	msg[6] = cpu_to_le32(0x00000000);
1198 	msg[7] = cpu_to_le32(0x00000780);
1199 	msg[8] = cpu_to_le32(0x00000440);
1200 	msg[9] = cpu_to_le32(0x00000000);
1201 	msg[10] = cpu_to_le32(0x01b37000);
1202 	for (i = 11; i < 1024; ++i)
1203 		msg[i] = cpu_to_le32(0x0);
1204 
1205 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1206 
1207 }
1208 
amdgpu_uvd_get_destroy_msg(struct amdgpu_ring * ring,uint32_t handle,bool direct,struct dma_fence ** fence)1209 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1210 			       bool direct, struct dma_fence **fence)
1211 {
1212 	struct amdgpu_device *adev = ring->adev;
1213 	struct amdgpu_bo *bo = NULL;
1214 	uint32_t *msg;
1215 	int r, i;
1216 
1217 	if (direct) {
1218 		bo = adev->uvd.ib_bo;
1219 	} else {
1220 		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1221 		if (r)
1222 			return r;
1223 	}
1224 
1225 	msg = amdgpu_bo_kptr(bo);
1226 	/* stitch together an UVD destroy msg */
1227 	msg[0] = cpu_to_le32(0x00000de4);
1228 	msg[1] = cpu_to_le32(0x00000002);
1229 	msg[2] = cpu_to_le32(handle);
1230 	msg[3] = cpu_to_le32(0x00000000);
1231 	for (i = 4; i < 1024; ++i)
1232 		msg[i] = cpu_to_le32(0x0);
1233 
1234 	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1235 
1236 	if (!direct)
1237 		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1238 
1239 	return r;
1240 }
1241 
amdgpu_uvd_idle_work_handler(struct work_struct * work)1242 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1243 {
1244 	struct amdgpu_device *adev =
1245 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1246 	unsigned int fences = 0, i, j;
1247 
1248 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1249 		if (adev->uvd.harvest_config & (1 << i))
1250 			continue;
1251 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1252 		for (j = 0; j < adev->uvd.num_enc_rings; ++j)
1253 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1254 	}
1255 
1256 	if (fences == 0) {
1257 		if (adev->pm.dpm_enabled) {
1258 			amdgpu_dpm_enable_uvd(adev, false);
1259 		} else {
1260 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1261 			/* shutdown the UVD block */
1262 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1263 							       AMD_PG_STATE_GATE);
1264 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1265 							       AMD_CG_STATE_GATE);
1266 		}
1267 	} else {
1268 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1269 	}
1270 }
1271 
amdgpu_uvd_ring_begin_use(struct amdgpu_ring * ring)1272 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1273 {
1274 	struct amdgpu_device *adev = ring->adev;
1275 	bool set_clocks;
1276 
1277 	if (amdgpu_sriov_vf(adev))
1278 		return;
1279 
1280 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1281 	if (set_clocks) {
1282 		if (adev->pm.dpm_enabled) {
1283 			amdgpu_dpm_enable_uvd(adev, true);
1284 		} else {
1285 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1286 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1287 							       AMD_CG_STATE_UNGATE);
1288 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1289 							       AMD_PG_STATE_UNGATE);
1290 		}
1291 	}
1292 }
1293 
amdgpu_uvd_ring_end_use(struct amdgpu_ring * ring)1294 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1295 {
1296 	if (!amdgpu_sriov_vf(ring->adev))
1297 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1298 }
1299 
1300 /**
1301  * amdgpu_uvd_ring_test_ib - test ib execution
1302  *
1303  * @ring: amdgpu_ring pointer
1304  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1305  *
1306  * Test if we can successfully execute an IB
1307  */
amdgpu_uvd_ring_test_ib(struct amdgpu_ring * ring,long timeout)1308 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1309 {
1310 	struct dma_fence *fence;
1311 	long r;
1312 
1313 	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1314 	if (r)
1315 		goto error;
1316 
1317 	r = dma_fence_wait_timeout(fence, false, timeout);
1318 	dma_fence_put(fence);
1319 	if (r == 0)
1320 		r = -ETIMEDOUT;
1321 	if (r < 0)
1322 		goto error;
1323 
1324 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1325 	if (r)
1326 		goto error;
1327 
1328 	r = dma_fence_wait_timeout(fence, false, timeout);
1329 	if (r == 0)
1330 		r = -ETIMEDOUT;
1331 	else if (r > 0)
1332 		r = 0;
1333 
1334 	dma_fence_put(fence);
1335 
1336 error:
1337 	return r;
1338 }
1339 
1340 /**
1341  * amdgpu_uvd_used_handles - returns used UVD handles
1342  *
1343  * @adev: amdgpu_device pointer
1344  *
1345  * Returns the number of UVD handles in use
1346  */
amdgpu_uvd_used_handles(struct amdgpu_device * adev)1347 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1348 {
1349 	unsigned int i;
1350 	uint32_t used_handles = 0;
1351 
1352 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1353 		/*
1354 		 * Handles can be freed in any order, and not
1355 		 * necessarily linear. So we need to count
1356 		 * all non-zero handles.
1357 		 */
1358 		if (atomic_read(&adev->uvd.handles[i]))
1359 			used_handles++;
1360 	}
1361 
1362 	return used_handles;
1363 }
1364