1 /*
2  * Copyright 2013 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  * Authors: Christian König <christian.koenig@amd.com>
26  */
27 
28 #include <linux/firmware.h>
29 #include <linux/module.h>
30 #include <drm/drmP.h>
31 #include <drm/drm.h>
32 
33 #include "amdgpu.h"
34 #include "amdgpu_pm.h"
35 #include "amdgpu_vce.h"
36 #include "cikd.h"
37 
38 /* 1 second timeout */
39 #define VCE_IDLE_TIMEOUT	msecs_to_jiffies(1000)
40 
41 /* Firmware Names */
42 #ifdef CONFIG_DRM_AMDGPU_CIK
43 #define FIRMWARE_BONAIRE	"radeon/bonaire_vce.bin"
44 #define FIRMWARE_KABINI	"radeon/kabini_vce.bin"
45 #define FIRMWARE_KAVERI	"radeon/kaveri_vce.bin"
46 #define FIRMWARE_HAWAII	"radeon/hawaii_vce.bin"
47 #define FIRMWARE_MULLINS	"radeon/mullins_vce.bin"
48 #endif
49 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
50 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
51 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
52 #define FIRMWARE_STONEY		"amdgpu/stoney_vce.bin"
53 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_vce.bin"
54 #define FIRMWARE_POLARIS11         "amdgpu/polaris11_vce.bin"
55 #define FIRMWARE_POLARIS12         "amdgpu/polaris12_vce.bin"
56 
57 #define FIRMWARE_VEGA10		"amdgpu/vega10_vce.bin"
58 
59 #ifdef CONFIG_DRM_AMDGPU_CIK
60 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
61 MODULE_FIRMWARE(FIRMWARE_KABINI);
62 MODULE_FIRMWARE(FIRMWARE_KAVERI);
63 MODULE_FIRMWARE(FIRMWARE_HAWAII);
64 MODULE_FIRMWARE(FIRMWARE_MULLINS);
65 #endif
66 MODULE_FIRMWARE(FIRMWARE_TONGA);
67 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
68 MODULE_FIRMWARE(FIRMWARE_FIJI);
69 MODULE_FIRMWARE(FIRMWARE_STONEY);
70 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
71 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
72 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
73 
74 MODULE_FIRMWARE(FIRMWARE_VEGA10);
75 
76 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
77 
78 /**
79  * amdgpu_vce_init - allocate memory, load vce firmware
80  *
81  * @adev: amdgpu_device pointer
82  *
83  * First step to get VCE online, allocate memory and load the firmware
84  */
85 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
86 {
87 	struct amdgpu_ring *ring;
88 	struct amd_sched_rq *rq;
89 	const char *fw_name;
90 	const struct common_firmware_header *hdr;
91 	unsigned ucode_version, version_major, version_minor, binary_id;
92 	int i, r;
93 
94 	switch (adev->asic_type) {
95 #ifdef CONFIG_DRM_AMDGPU_CIK
96 	case CHIP_BONAIRE:
97 		fw_name = FIRMWARE_BONAIRE;
98 		break;
99 	case CHIP_KAVERI:
100 		fw_name = FIRMWARE_KAVERI;
101 		break;
102 	case CHIP_KABINI:
103 		fw_name = FIRMWARE_KABINI;
104 		break;
105 	case CHIP_HAWAII:
106 		fw_name = FIRMWARE_HAWAII;
107 		break;
108 	case CHIP_MULLINS:
109 		fw_name = FIRMWARE_MULLINS;
110 		break;
111 #endif
112 	case CHIP_TONGA:
113 		fw_name = FIRMWARE_TONGA;
114 		break;
115 	case CHIP_CARRIZO:
116 		fw_name = FIRMWARE_CARRIZO;
117 		break;
118 	case CHIP_FIJI:
119 		fw_name = FIRMWARE_FIJI;
120 		break;
121 	case CHIP_STONEY:
122 		fw_name = FIRMWARE_STONEY;
123 		break;
124 	case CHIP_POLARIS10:
125 		fw_name = FIRMWARE_POLARIS10;
126 		break;
127 	case CHIP_POLARIS11:
128 		fw_name = FIRMWARE_POLARIS11;
129 		break;
130 	case CHIP_VEGA10:
131 		fw_name = FIRMWARE_VEGA10;
132 		break;
133 	case CHIP_POLARIS12:
134 		fw_name = FIRMWARE_POLARIS12;
135 		break;
136 
137 	default:
138 		return -EINVAL;
139 	}
140 
141 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
142 	if (r) {
143 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
144 			fw_name);
145 		return r;
146 	}
147 
148 	r = amdgpu_ucode_validate(adev->vce.fw);
149 	if (r) {
150 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
151 			fw_name);
152 		release_firmware(adev->vce.fw);
153 		adev->vce.fw = NULL;
154 		return r;
155 	}
156 
157 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
158 
159 	ucode_version = le32_to_cpu(hdr->ucode_version);
160 	version_major = (ucode_version >> 20) & 0xfff;
161 	version_minor = (ucode_version >> 8) & 0xfff;
162 	binary_id = ucode_version & 0xff;
163 	DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
164 		version_major, version_minor, binary_id);
165 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
166 				(binary_id << 8));
167 
168 	/* allocate firmware, stack and heap BO */
169 
170 	r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
171 			     AMDGPU_GEM_DOMAIN_VRAM,
172 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
173 			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
174 			     NULL, NULL, &adev->vce.vcpu_bo);
175 	if (r) {
176 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
177 		return r;
178 	}
179 
180 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
181 	if (r) {
182 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
183 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
184 		return r;
185 	}
186 
187 	r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
188 			  &adev->vce.gpu_addr);
189 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
190 	if (r) {
191 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
192 		dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
193 		return r;
194 	}
195 
196 
197 	ring = &adev->vce.ring[0];
198 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
199 	r = amd_sched_entity_init(&ring->sched, &adev->vce.entity,
200 				  rq, amdgpu_sched_jobs);
201 	if (r != 0) {
202 		DRM_ERROR("Failed setting up VCE run queue.\n");
203 		return r;
204 	}
205 
206 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
207 		atomic_set(&adev->vce.handles[i], 0);
208 		adev->vce.filp[i] = NULL;
209 	}
210 
211 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
212 	mutex_init(&adev->vce.idle_mutex);
213 
214 	return 0;
215 }
216 
217 /**
218  * amdgpu_vce_fini - free memory
219  *
220  * @adev: amdgpu_device pointer
221  *
222  * Last step on VCE teardown, free firmware memory
223  */
224 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
225 {
226 	unsigned i;
227 
228 	if (adev->vce.vcpu_bo == NULL)
229 		return 0;
230 
231 	amd_sched_entity_fini(&adev->vce.ring[0].sched, &adev->vce.entity);
232 
233 	amdgpu_bo_unref(&adev->vce.vcpu_bo);
234 
235 	for (i = 0; i < adev->vce.num_rings; i++)
236 		amdgpu_ring_fini(&adev->vce.ring[i]);
237 
238 	release_firmware(adev->vce.fw);
239 	mutex_destroy(&adev->vce.idle_mutex);
240 
241 	return 0;
242 }
243 
244 /**
245  * amdgpu_vce_suspend - unpin VCE fw memory
246  *
247  * @adev: amdgpu_device pointer
248  *
249  */
250 int amdgpu_vce_suspend(struct amdgpu_device *adev)
251 {
252 	int i;
253 
254 	if (adev->vce.vcpu_bo == NULL)
255 		return 0;
256 
257 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
258 		if (atomic_read(&adev->vce.handles[i]))
259 			break;
260 
261 	if (i == AMDGPU_MAX_VCE_HANDLES)
262 		return 0;
263 
264 	cancel_delayed_work_sync(&adev->vce.idle_work);
265 	/* TODO: suspending running encoding sessions isn't supported */
266 	return -EINVAL;
267 }
268 
269 /**
270  * amdgpu_vce_resume - pin VCE fw memory
271  *
272  * @adev: amdgpu_device pointer
273  *
274  */
275 int amdgpu_vce_resume(struct amdgpu_device *adev)
276 {
277 	void *cpu_addr;
278 	const struct common_firmware_header *hdr;
279 	unsigned offset;
280 	int r;
281 
282 	if (adev->vce.vcpu_bo == NULL)
283 		return -EINVAL;
284 
285 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
286 	if (r) {
287 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
288 		return r;
289 	}
290 
291 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
292 	if (r) {
293 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
294 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
295 		return r;
296 	}
297 
298 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
299 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
300 	memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
301 		    adev->vce.fw->size - offset);
302 
303 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
304 
305 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
306 
307 	return 0;
308 }
309 
310 /**
311  * amdgpu_vce_idle_work_handler - power off VCE
312  *
313  * @work: pointer to work structure
314  *
315  * power of VCE when it's not used any more
316  */
317 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
318 {
319 	struct amdgpu_device *adev =
320 		container_of(work, struct amdgpu_device, vce.idle_work.work);
321 	unsigned i, count = 0;
322 
323 	if (amdgpu_sriov_vf(adev))
324 		return;
325 
326 	for (i = 0; i < adev->vce.num_rings; i++)
327 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
328 
329 	if (count == 0) {
330 		if (adev->pm.dpm_enabled) {
331 			amdgpu_dpm_enable_vce(adev, false);
332 		} else {
333 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
334 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
335 							    AMD_PG_STATE_GATE);
336 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
337 							    AMD_CG_STATE_GATE);
338 		}
339 	} else {
340 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
341 	}
342 }
343 
344 /**
345  * amdgpu_vce_ring_begin_use - power up VCE
346  *
347  * @ring: amdgpu ring
348  *
349  * Make sure VCE is powerd up when we want to use it
350  */
351 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
352 {
353 	struct amdgpu_device *adev = ring->adev;
354 	bool set_clocks;
355 
356 	if (amdgpu_sriov_vf(adev))
357 		return;
358 
359 	mutex_lock(&adev->vce.idle_mutex);
360 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
361 	if (set_clocks) {
362 		if (adev->pm.dpm_enabled) {
363 			amdgpu_dpm_enable_vce(adev, true);
364 		} else {
365 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
366 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
367 							    AMD_CG_STATE_UNGATE);
368 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
369 							    AMD_PG_STATE_UNGATE);
370 
371 		}
372 	}
373 	mutex_unlock(&adev->vce.idle_mutex);
374 }
375 
376 /**
377  * amdgpu_vce_ring_end_use - power VCE down
378  *
379  * @ring: amdgpu ring
380  *
381  * Schedule work to power VCE down again
382  */
383 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
384 {
385 	schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
386 }
387 
388 /**
389  * amdgpu_vce_free_handles - free still open VCE handles
390  *
391  * @adev: amdgpu_device pointer
392  * @filp: drm file pointer
393  *
394  * Close all VCE handles still open by this file pointer
395  */
396 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
397 {
398 	struct amdgpu_ring *ring = &adev->vce.ring[0];
399 	int i, r;
400 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
401 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
402 
403 		if (!handle || adev->vce.filp[i] != filp)
404 			continue;
405 
406 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
407 		if (r)
408 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
409 
410 		adev->vce.filp[i] = NULL;
411 		atomic_set(&adev->vce.handles[i], 0);
412 	}
413 }
414 
415 /**
416  * amdgpu_vce_get_create_msg - generate a VCE create msg
417  *
418  * @adev: amdgpu_device pointer
419  * @ring: ring we should submit the msg to
420  * @handle: VCE session handle to use
421  * @fence: optional fence to return
422  *
423  * Open up a stream for HW test
424  */
425 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
426 			      struct dma_fence **fence)
427 {
428 	const unsigned ib_size_dw = 1024;
429 	struct amdgpu_job *job;
430 	struct amdgpu_ib *ib;
431 	struct dma_fence *f = NULL;
432 	uint64_t dummy;
433 	int i, r;
434 
435 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
436 	if (r)
437 		return r;
438 
439 	ib = &job->ibs[0];
440 
441 	dummy = ib->gpu_addr + 1024;
442 
443 	/* stitch together an VCE create msg */
444 	ib->length_dw = 0;
445 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
446 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
447 	ib->ptr[ib->length_dw++] = handle;
448 
449 	if ((ring->adev->vce.fw_version >> 24) >= 52)
450 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
451 	else
452 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
453 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
454 	ib->ptr[ib->length_dw++] = 0x00000000;
455 	ib->ptr[ib->length_dw++] = 0x00000042;
456 	ib->ptr[ib->length_dw++] = 0x0000000a;
457 	ib->ptr[ib->length_dw++] = 0x00000001;
458 	ib->ptr[ib->length_dw++] = 0x00000080;
459 	ib->ptr[ib->length_dw++] = 0x00000060;
460 	ib->ptr[ib->length_dw++] = 0x00000100;
461 	ib->ptr[ib->length_dw++] = 0x00000100;
462 	ib->ptr[ib->length_dw++] = 0x0000000c;
463 	ib->ptr[ib->length_dw++] = 0x00000000;
464 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
465 		ib->ptr[ib->length_dw++] = 0x00000000;
466 		ib->ptr[ib->length_dw++] = 0x00000000;
467 		ib->ptr[ib->length_dw++] = 0x00000000;
468 		ib->ptr[ib->length_dw++] = 0x00000000;
469 	}
470 
471 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
472 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
473 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
474 	ib->ptr[ib->length_dw++] = dummy;
475 	ib->ptr[ib->length_dw++] = 0x00000001;
476 
477 	for (i = ib->length_dw; i < ib_size_dw; ++i)
478 		ib->ptr[i] = 0x0;
479 
480 	r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
481 	job->fence = dma_fence_get(f);
482 	if (r)
483 		goto err;
484 
485 	amdgpu_job_free(job);
486 	if (fence)
487 		*fence = dma_fence_get(f);
488 	dma_fence_put(f);
489 	return 0;
490 
491 err:
492 	amdgpu_job_free(job);
493 	return r;
494 }
495 
496 /**
497  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
498  *
499  * @adev: amdgpu_device pointer
500  * @ring: ring we should submit the msg to
501  * @handle: VCE session handle to use
502  * @fence: optional fence to return
503  *
504  * Close up a stream for HW test or if userspace failed to do so
505  */
506 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
507 			       bool direct, struct dma_fence **fence)
508 {
509 	const unsigned ib_size_dw = 1024;
510 	struct amdgpu_job *job;
511 	struct amdgpu_ib *ib;
512 	struct dma_fence *f = NULL;
513 	int i, r;
514 
515 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
516 	if (r)
517 		return r;
518 
519 	ib = &job->ibs[0];
520 
521 	/* stitch together an VCE destroy msg */
522 	ib->length_dw = 0;
523 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
524 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
525 	ib->ptr[ib->length_dw++] = handle;
526 
527 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
528 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
529 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
530 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
531 	ib->ptr[ib->length_dw++] = 0x00000000;
532 	ib->ptr[ib->length_dw++] = 0x00000000;
533 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
534 	ib->ptr[ib->length_dw++] = 0x00000000;
535 
536 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
537 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
538 
539 	for (i = ib->length_dw; i < ib_size_dw; ++i)
540 		ib->ptr[i] = 0x0;
541 
542 	if (direct) {
543 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
544 		job->fence = dma_fence_get(f);
545 		if (r)
546 			goto err;
547 
548 		amdgpu_job_free(job);
549 	} else {
550 		r = amdgpu_job_submit(job, ring, &ring->adev->vce.entity,
551 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
552 		if (r)
553 			goto err;
554 	}
555 
556 	if (fence)
557 		*fence = dma_fence_get(f);
558 	dma_fence_put(f);
559 	return 0;
560 
561 err:
562 	amdgpu_job_free(job);
563 	return r;
564 }
565 
566 /**
567  * amdgpu_vce_cs_reloc - command submission relocation
568  *
569  * @p: parser context
570  * @lo: address of lower dword
571  * @hi: address of higher dword
572  * @size: minimum size
573  *
574  * Patch relocation inside command stream with real buffer address
575  */
576 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
577 			       int lo, int hi, unsigned size, uint32_t index)
578 {
579 	struct amdgpu_bo_va_mapping *mapping;
580 	struct amdgpu_bo *bo;
581 	uint64_t addr;
582 
583 	if (index == 0xffffffff)
584 		index = 0;
585 
586 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
587 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
588 	addr += ((uint64_t)size) * ((uint64_t)index);
589 
590 	mapping = amdgpu_cs_find_mapping(p, addr, &bo);
591 	if (mapping == NULL) {
592 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
593 			  addr, lo, hi, size, index);
594 		return -EINVAL;
595 	}
596 
597 	if ((addr + (uint64_t)size) >
598 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
599 		DRM_ERROR("BO to small for addr 0x%010Lx %d %d\n",
600 			  addr, lo, hi);
601 		return -EINVAL;
602 	}
603 
604 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
605 	addr += amdgpu_bo_gpu_offset(bo);
606 	addr -= ((uint64_t)size) * ((uint64_t)index);
607 
608 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
609 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
610 
611 	return 0;
612 }
613 
614 /**
615  * amdgpu_vce_validate_handle - validate stream handle
616  *
617  * @p: parser context
618  * @handle: handle to validate
619  * @allocated: allocated a new handle?
620  *
621  * Validates the handle and return the found session index or -EINVAL
622  * we we don't have another free session index.
623  */
624 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
625 				      uint32_t handle, uint32_t *allocated)
626 {
627 	unsigned i;
628 
629 	/* validate the handle */
630 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
631 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
632 			if (p->adev->vce.filp[i] != p->filp) {
633 				DRM_ERROR("VCE handle collision detected!\n");
634 				return -EINVAL;
635 			}
636 			return i;
637 		}
638 	}
639 
640 	/* handle not found try to alloc a new one */
641 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
642 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
643 			p->adev->vce.filp[i] = p->filp;
644 			p->adev->vce.img_size[i] = 0;
645 			*allocated |= 1 << i;
646 			return i;
647 		}
648 	}
649 
650 	DRM_ERROR("No more free VCE handles!\n");
651 	return -EINVAL;
652 }
653 
654 /**
655  * amdgpu_vce_cs_parse - parse and validate the command stream
656  *
657  * @p: parser context
658  *
659  */
660 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
661 {
662 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
663 	unsigned fb_idx = 0, bs_idx = 0;
664 	int session_idx = -1;
665 	uint32_t destroyed = 0;
666 	uint32_t created = 0;
667 	uint32_t allocated = 0;
668 	uint32_t tmp, handle = 0;
669 	uint32_t *size = &tmp;
670 	int i, r, idx = 0;
671 
672 	p->job->vm = NULL;
673 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
674 
675 	r = amdgpu_cs_sysvm_access_required(p);
676 	if (r)
677 		return r;
678 
679 	while (idx < ib->length_dw) {
680 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
681 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
682 
683 		if ((len < 8) || (len & 3)) {
684 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
685 			r = -EINVAL;
686 			goto out;
687 		}
688 
689 		switch (cmd) {
690 		case 0x00000001: /* session */
691 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
692 			session_idx = amdgpu_vce_validate_handle(p, handle,
693 								 &allocated);
694 			if (session_idx < 0) {
695 				r = session_idx;
696 				goto out;
697 			}
698 			size = &p->adev->vce.img_size[session_idx];
699 			break;
700 
701 		case 0x00000002: /* task info */
702 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
703 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
704 			break;
705 
706 		case 0x01000001: /* create */
707 			created |= 1 << session_idx;
708 			if (destroyed & (1 << session_idx)) {
709 				destroyed &= ~(1 << session_idx);
710 				allocated |= 1 << session_idx;
711 
712 			} else if (!(allocated & (1 << session_idx))) {
713 				DRM_ERROR("Handle already in use!\n");
714 				r = -EINVAL;
715 				goto out;
716 			}
717 
718 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
719 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
720 				8 * 3 / 2;
721 			break;
722 
723 		case 0x04000001: /* config extension */
724 		case 0x04000002: /* pic control */
725 		case 0x04000005: /* rate control */
726 		case 0x04000007: /* motion estimation */
727 		case 0x04000008: /* rdo */
728 		case 0x04000009: /* vui */
729 		case 0x05000002: /* auxiliary buffer */
730 		case 0x05000009: /* clock table */
731 			break;
732 
733 		case 0x0500000c: /* hw config */
734 			switch (p->adev->asic_type) {
735 #ifdef CONFIG_DRM_AMDGPU_CIK
736 			case CHIP_KAVERI:
737 			case CHIP_MULLINS:
738 #endif
739 			case CHIP_CARRIZO:
740 				break;
741 			default:
742 				r = -EINVAL;
743 				goto out;
744 			}
745 			break;
746 
747 		case 0x03000001: /* encode */
748 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
749 						*size, 0);
750 			if (r)
751 				goto out;
752 
753 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
754 						*size / 3, 0);
755 			if (r)
756 				goto out;
757 			break;
758 
759 		case 0x02000001: /* destroy */
760 			destroyed |= 1 << session_idx;
761 			break;
762 
763 		case 0x05000001: /* context buffer */
764 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
765 						*size * 2, 0);
766 			if (r)
767 				goto out;
768 			break;
769 
770 		case 0x05000004: /* video bitstream buffer */
771 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
772 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
773 						tmp, bs_idx);
774 			if (r)
775 				goto out;
776 			break;
777 
778 		case 0x05000005: /* feedback buffer */
779 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
780 						4096, fb_idx);
781 			if (r)
782 				goto out;
783 			break;
784 
785 		default:
786 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
787 			r = -EINVAL;
788 			goto out;
789 		}
790 
791 		if (session_idx == -1) {
792 			DRM_ERROR("no session command at start of IB\n");
793 			r = -EINVAL;
794 			goto out;
795 		}
796 
797 		idx += len / 4;
798 	}
799 
800 	if (allocated & ~created) {
801 		DRM_ERROR("New session without create command!\n");
802 		r = -ENOENT;
803 	}
804 
805 out:
806 	if (!r) {
807 		/* No error, free all destroyed handle slots */
808 		tmp = destroyed;
809 	} else {
810 		/* Error during parsing, free all allocated handle slots */
811 		tmp = allocated;
812 	}
813 
814 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
815 		if (tmp & (1 << i))
816 			atomic_set(&p->adev->vce.handles[i], 0);
817 
818 	return r;
819 }
820 
821 /**
822  * amdgpu_vce_cs_parse_vm - parse the command stream in VM mode
823  *
824  * @p: parser context
825  *
826  */
827 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p, uint32_t ib_idx)
828 {
829 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
830 	int session_idx = -1;
831 	uint32_t destroyed = 0;
832 	uint32_t created = 0;
833 	uint32_t allocated = 0;
834 	uint32_t tmp, handle = 0;
835 	int i, r = 0, idx = 0;
836 
837 	while (idx < ib->length_dw) {
838 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
839 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
840 
841 		if ((len < 8) || (len & 3)) {
842 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
843 			r = -EINVAL;
844 			goto out;
845 		}
846 
847 		switch (cmd) {
848 		case 0x00000001: /* session */
849 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
850 			session_idx = amdgpu_vce_validate_handle(p, handle,
851 								 &allocated);
852 			if (session_idx < 0) {
853 				r = session_idx;
854 				goto out;
855 			}
856 			break;
857 
858 		case 0x01000001: /* create */
859 			created |= 1 << session_idx;
860 			if (destroyed & (1 << session_idx)) {
861 				destroyed &= ~(1 << session_idx);
862 				allocated |= 1 << session_idx;
863 
864 			} else if (!(allocated & (1 << session_idx))) {
865 				DRM_ERROR("Handle already in use!\n");
866 				r = -EINVAL;
867 				goto out;
868 			}
869 
870 			break;
871 
872 		case 0x02000001: /* destroy */
873 			destroyed |= 1 << session_idx;
874 			break;
875 
876 		default:
877 			break;
878 		}
879 
880 		if (session_idx == -1) {
881 			DRM_ERROR("no session command at start of IB\n");
882 			r = -EINVAL;
883 			goto out;
884 		}
885 
886 		idx += len / 4;
887 	}
888 
889 	if (allocated & ~created) {
890 		DRM_ERROR("New session without create command!\n");
891 		r = -ENOENT;
892 	}
893 
894 out:
895 	if (!r) {
896 		/* No error, free all destroyed handle slots */
897 		tmp = destroyed;
898 		amdgpu_ib_free(p->adev, ib, NULL);
899 	} else {
900 		/* Error during parsing, free all allocated handle slots */
901 		tmp = allocated;
902 	}
903 
904 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
905 		if (tmp & (1 << i))
906 			atomic_set(&p->adev->vce.handles[i], 0);
907 
908 	return r;
909 }
910 
911 /**
912  * amdgpu_vce_ring_emit_ib - execute indirect buffer
913  *
914  * @ring: engine to use
915  * @ib: the IB to execute
916  *
917  */
918 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib,
919 			     unsigned vm_id, bool ctx_switch)
920 {
921 	amdgpu_ring_write(ring, VCE_CMD_IB);
922 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
923 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
924 	amdgpu_ring_write(ring, ib->length_dw);
925 }
926 
927 /**
928  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
929  *
930  * @ring: engine to use
931  * @fence: the fence
932  *
933  */
934 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
935 				unsigned flags)
936 {
937 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
938 
939 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
940 	amdgpu_ring_write(ring, addr);
941 	amdgpu_ring_write(ring, upper_32_bits(addr));
942 	amdgpu_ring_write(ring, seq);
943 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
944 	amdgpu_ring_write(ring, VCE_CMD_END);
945 }
946 
947 /**
948  * amdgpu_vce_ring_test_ring - test if VCE ring is working
949  *
950  * @ring: the engine to test on
951  *
952  */
953 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
954 {
955 	struct amdgpu_device *adev = ring->adev;
956 	uint32_t rptr = amdgpu_ring_get_rptr(ring);
957 	unsigned i;
958 	int r, timeout = adev->usec_timeout;
959 
960 	/* workaround VCE ring test slow issue for sriov*/
961 	if (amdgpu_sriov_vf(adev))
962 		timeout *= 10;
963 
964 	r = amdgpu_ring_alloc(ring, 16);
965 	if (r) {
966 		DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
967 			  ring->idx, r);
968 		return r;
969 	}
970 	amdgpu_ring_write(ring, VCE_CMD_END);
971 	amdgpu_ring_commit(ring);
972 
973 	for (i = 0; i < timeout; i++) {
974 		if (amdgpu_ring_get_rptr(ring) != rptr)
975 			break;
976 		DRM_UDELAY(1);
977 	}
978 
979 	if (i < timeout) {
980 		DRM_INFO("ring test on %d succeeded in %d usecs\n",
981 			 ring->idx, i);
982 	} else {
983 		DRM_ERROR("amdgpu: ring %d test failed\n",
984 			  ring->idx);
985 		r = -ETIMEDOUT;
986 	}
987 
988 	return r;
989 }
990 
991 /**
992  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
993  *
994  * @ring: the engine to test on
995  *
996  */
997 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
998 {
999 	struct dma_fence *fence = NULL;
1000 	long r;
1001 
1002 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1003 	if (ring != &ring->adev->vce.ring[0])
1004 		return 0;
1005 
1006 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
1007 	if (r) {
1008 		DRM_ERROR("amdgpu: failed to get create msg (%ld).\n", r);
1009 		goto error;
1010 	}
1011 
1012 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1013 	if (r) {
1014 		DRM_ERROR("amdgpu: failed to get destroy ib (%ld).\n", r);
1015 		goto error;
1016 	}
1017 
1018 	r = dma_fence_wait_timeout(fence, false, timeout);
1019 	if (r == 0) {
1020 		DRM_ERROR("amdgpu: IB test timed out.\n");
1021 		r = -ETIMEDOUT;
1022 	} else if (r < 0) {
1023 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1024 	} else {
1025 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
1026 		r = 0;
1027 	}
1028 error:
1029 	dma_fence_put(fence);
1030 	return r;
1031 }
1032