xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c (revision 781095f903f398148cd0b646d3984234a715f29e)
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_MS	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 
54 #ifdef CONFIG_DRM_AMDGPU_CIK
55 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
56 MODULE_FIRMWARE(FIRMWARE_KABINI);
57 MODULE_FIRMWARE(FIRMWARE_KAVERI);
58 MODULE_FIRMWARE(FIRMWARE_HAWAII);
59 MODULE_FIRMWARE(FIRMWARE_MULLINS);
60 #endif
61 MODULE_FIRMWARE(FIRMWARE_TONGA);
62 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
63 MODULE_FIRMWARE(FIRMWARE_FIJI);
64 MODULE_FIRMWARE(FIRMWARE_STONEY);
65 
66 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
67 
68 /**
69  * amdgpu_vce_init - allocate memory, load vce firmware
70  *
71  * @adev: amdgpu_device pointer
72  *
73  * First step to get VCE online, allocate memory and load the firmware
74  */
75 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
76 {
77 	const char *fw_name;
78 	const struct common_firmware_header *hdr;
79 	unsigned ucode_version, version_major, version_minor, binary_id;
80 	int i, r;
81 
82 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
83 
84 	switch (adev->asic_type) {
85 #ifdef CONFIG_DRM_AMDGPU_CIK
86 	case CHIP_BONAIRE:
87 		fw_name = FIRMWARE_BONAIRE;
88 		break;
89 	case CHIP_KAVERI:
90 		fw_name = FIRMWARE_KAVERI;
91 		break;
92 	case CHIP_KABINI:
93 		fw_name = FIRMWARE_KABINI;
94 		break;
95 	case CHIP_HAWAII:
96 		fw_name = FIRMWARE_HAWAII;
97 		break;
98 	case CHIP_MULLINS:
99 		fw_name = FIRMWARE_MULLINS;
100 		break;
101 #endif
102 	case CHIP_TONGA:
103 		fw_name = FIRMWARE_TONGA;
104 		break;
105 	case CHIP_CARRIZO:
106 		fw_name = FIRMWARE_CARRIZO;
107 		break;
108 	case CHIP_FIJI:
109 		fw_name = FIRMWARE_FIJI;
110 		break;
111 	case CHIP_STONEY:
112 		fw_name = FIRMWARE_STONEY;
113 		break;
114 
115 	default:
116 		return -EINVAL;
117 	}
118 
119 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
120 	if (r) {
121 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
122 			fw_name);
123 		return r;
124 	}
125 
126 	r = amdgpu_ucode_validate(adev->vce.fw);
127 	if (r) {
128 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
129 			fw_name);
130 		release_firmware(adev->vce.fw);
131 		adev->vce.fw = NULL;
132 		return r;
133 	}
134 
135 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
136 
137 	ucode_version = le32_to_cpu(hdr->ucode_version);
138 	version_major = (ucode_version >> 20) & 0xfff;
139 	version_minor = (ucode_version >> 8) & 0xfff;
140 	binary_id = ucode_version & 0xff;
141 	DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
142 		version_major, version_minor, binary_id);
143 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
144 				(binary_id << 8));
145 
146 	/* allocate firmware, stack and heap BO */
147 
148 	r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
149 			     AMDGPU_GEM_DOMAIN_VRAM,
150 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
151 			     NULL, NULL, &adev->vce.vcpu_bo);
152 	if (r) {
153 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
154 		return r;
155 	}
156 
157 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
158 	if (r) {
159 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
160 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
161 		return r;
162 	}
163 
164 	r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
165 			  &adev->vce.gpu_addr);
166 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
167 	if (r) {
168 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
169 		dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
170 		return r;
171 	}
172 
173 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
174 		atomic_set(&adev->vce.handles[i], 0);
175 		adev->vce.filp[i] = NULL;
176 	}
177 
178 	return 0;
179 }
180 
181 /**
182  * amdgpu_vce_fini - free memory
183  *
184  * @adev: amdgpu_device pointer
185  *
186  * Last step on VCE teardown, free firmware memory
187  */
188 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
189 {
190 	if (adev->vce.vcpu_bo == NULL)
191 		return 0;
192 
193 	amdgpu_bo_unref(&adev->vce.vcpu_bo);
194 
195 	amdgpu_ring_fini(&adev->vce.ring[0]);
196 	amdgpu_ring_fini(&adev->vce.ring[1]);
197 
198 	release_firmware(adev->vce.fw);
199 
200 	return 0;
201 }
202 
203 /**
204  * amdgpu_vce_suspend - unpin VCE fw memory
205  *
206  * @adev: amdgpu_device pointer
207  *
208  */
209 int amdgpu_vce_suspend(struct amdgpu_device *adev)
210 {
211 	int i;
212 
213 	if (adev->vce.vcpu_bo == NULL)
214 		return 0;
215 
216 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
217 		if (atomic_read(&adev->vce.handles[i]))
218 			break;
219 
220 	if (i == AMDGPU_MAX_VCE_HANDLES)
221 		return 0;
222 
223 	/* TODO: suspending running encoding sessions isn't supported */
224 	return -EINVAL;
225 }
226 
227 /**
228  * amdgpu_vce_resume - pin VCE fw memory
229  *
230  * @adev: amdgpu_device pointer
231  *
232  */
233 int amdgpu_vce_resume(struct amdgpu_device *adev)
234 {
235 	void *cpu_addr;
236 	const struct common_firmware_header *hdr;
237 	unsigned offset;
238 	int r;
239 
240 	if (adev->vce.vcpu_bo == NULL)
241 		return -EINVAL;
242 
243 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
244 	if (r) {
245 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
246 		return r;
247 	}
248 
249 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
250 	if (r) {
251 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
252 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
253 		return r;
254 	}
255 
256 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
257 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
258 	memcpy(cpu_addr, (adev->vce.fw->data) + offset,
259 		(adev->vce.fw->size) - offset);
260 
261 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
262 
263 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
264 
265 	return 0;
266 }
267 
268 /**
269  * amdgpu_vce_idle_work_handler - power off VCE
270  *
271  * @work: pointer to work structure
272  *
273  * power of VCE when it's not used any more
274  */
275 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
276 {
277 	struct amdgpu_device *adev =
278 		container_of(work, struct amdgpu_device, vce.idle_work.work);
279 
280 	if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
281 	    (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
282 		if (adev->pm.dpm_enabled) {
283 			amdgpu_dpm_enable_vce(adev, false);
284 		} else {
285 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
286 		}
287 	} else {
288 		schedule_delayed_work(&adev->vce.idle_work,
289 				      msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
290 	}
291 }
292 
293 /**
294  * amdgpu_vce_note_usage - power up VCE
295  *
296  * @adev: amdgpu_device pointer
297  *
298  * Make sure VCE is powerd up when we want to use it
299  */
300 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
301 {
302 	bool streams_changed = false;
303 	bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
304 	set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
305 					    msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
306 
307 	if (adev->pm.dpm_enabled) {
308 		/* XXX figure out if the streams changed */
309 		streams_changed = false;
310 	}
311 
312 	if (set_clocks || streams_changed) {
313 		if (adev->pm.dpm_enabled) {
314 			amdgpu_dpm_enable_vce(adev, true);
315 		} else {
316 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
317 		}
318 	}
319 }
320 
321 /**
322  * amdgpu_vce_free_handles - free still open VCE handles
323  *
324  * @adev: amdgpu_device pointer
325  * @filp: drm file pointer
326  *
327  * Close all VCE handles still open by this file pointer
328  */
329 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
330 {
331 	struct amdgpu_ring *ring = &adev->vce.ring[0];
332 	int i, r;
333 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
334 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
335 		if (!handle || adev->vce.filp[i] != filp)
336 			continue;
337 
338 		amdgpu_vce_note_usage(adev);
339 
340 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
341 		if (r)
342 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
343 
344 		adev->vce.filp[i] = NULL;
345 		atomic_set(&adev->vce.handles[i], 0);
346 	}
347 }
348 
349 /**
350  * amdgpu_vce_get_create_msg - generate a VCE create msg
351  *
352  * @adev: amdgpu_device pointer
353  * @ring: ring we should submit the msg to
354  * @handle: VCE session handle to use
355  * @fence: optional fence to return
356  *
357  * Open up a stream for HW test
358  */
359 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
360 			      struct fence **fence)
361 {
362 	const unsigned ib_size_dw = 1024;
363 	struct amdgpu_job *job;
364 	struct amdgpu_ib *ib;
365 	struct fence *f = NULL;
366 	uint64_t dummy;
367 	int i, r;
368 
369 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
370 	if (r)
371 		return r;
372 
373 	ib = &job->ibs[0];
374 
375 	dummy = ib->gpu_addr + 1024;
376 
377 	/* stitch together an VCE create msg */
378 	ib->length_dw = 0;
379 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
380 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
381 	ib->ptr[ib->length_dw++] = handle;
382 
383 	if ((ring->adev->vce.fw_version >> 24) >= 52)
384 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
385 	else
386 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
387 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
388 	ib->ptr[ib->length_dw++] = 0x00000000;
389 	ib->ptr[ib->length_dw++] = 0x00000042;
390 	ib->ptr[ib->length_dw++] = 0x0000000a;
391 	ib->ptr[ib->length_dw++] = 0x00000001;
392 	ib->ptr[ib->length_dw++] = 0x00000080;
393 	ib->ptr[ib->length_dw++] = 0x00000060;
394 	ib->ptr[ib->length_dw++] = 0x00000100;
395 	ib->ptr[ib->length_dw++] = 0x00000100;
396 	ib->ptr[ib->length_dw++] = 0x0000000c;
397 	ib->ptr[ib->length_dw++] = 0x00000000;
398 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
399 		ib->ptr[ib->length_dw++] = 0x00000000;
400 		ib->ptr[ib->length_dw++] = 0x00000000;
401 		ib->ptr[ib->length_dw++] = 0x00000000;
402 		ib->ptr[ib->length_dw++] = 0x00000000;
403 	}
404 
405 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
406 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
407 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
408 	ib->ptr[ib->length_dw++] = dummy;
409 	ib->ptr[ib->length_dw++] = 0x00000001;
410 
411 	for (i = ib->length_dw; i < ib_size_dw; ++i)
412 		ib->ptr[i] = 0x0;
413 
414 	r = amdgpu_ib_schedule(ring, 1, ib, AMDGPU_FENCE_OWNER_UNDEFINED,
415 			       NULL, &f);
416 	if (r)
417 		goto err;
418 
419 	amdgpu_job_free(job);
420 	if (fence)
421 		*fence = fence_get(f);
422 	fence_put(f);
423 	return 0;
424 
425 err:
426 	amdgpu_job_free(job);
427 	return r;
428 }
429 
430 /**
431  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
432  *
433  * @adev: amdgpu_device pointer
434  * @ring: ring we should submit the msg to
435  * @handle: VCE session handle to use
436  * @fence: optional fence to return
437  *
438  * Close up a stream for HW test or if userspace failed to do so
439  */
440 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
441 			       bool direct, struct fence **fence)
442 {
443 	const unsigned ib_size_dw = 1024;
444 	struct amdgpu_job *job;
445 	struct amdgpu_ib *ib;
446 	struct fence *f = NULL;
447 	uint64_t dummy;
448 	int i, r;
449 
450 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
451 	if (r)
452 		return r;
453 
454 	ib = &job->ibs[0];
455 	dummy = ib->gpu_addr + 1024;
456 
457 	/* stitch together an VCE destroy msg */
458 	ib->length_dw = 0;
459 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
460 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
461 	ib->ptr[ib->length_dw++] = handle;
462 
463 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
464 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
465 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
466 	ib->ptr[ib->length_dw++] = dummy;
467 	ib->ptr[ib->length_dw++] = 0x00000001;
468 
469 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
470 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
471 
472 	for (i = ib->length_dw; i < ib_size_dw; ++i)
473 		ib->ptr[i] = 0x0;
474 
475 	if (direct) {
476 		r = amdgpu_ib_schedule(ring, 1, ib,
477 				       AMDGPU_FENCE_OWNER_UNDEFINED,
478 				       NULL, &f);
479 		if (r)
480 			goto err;
481 
482 		amdgpu_job_free(job);
483 	} else {
484 		r = amdgpu_job_submit(job, ring,
485 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
486 		if (r)
487 			goto err;
488 	}
489 
490 	if (fence)
491 		*fence = fence_get(f);
492 	fence_put(f);
493 	return 0;
494 
495 err:
496 	amdgpu_job_free(job);
497 	return r;
498 }
499 
500 /**
501  * amdgpu_vce_cs_reloc - command submission relocation
502  *
503  * @p: parser context
504  * @lo: address of lower dword
505  * @hi: address of higher dword
506  * @size: minimum size
507  *
508  * Patch relocation inside command stream with real buffer address
509  */
510 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
511 			       int lo, int hi, unsigned size, uint32_t index)
512 {
513 	struct amdgpu_bo_va_mapping *mapping;
514 	struct amdgpu_bo *bo;
515 	uint64_t addr;
516 
517 	if (index == 0xffffffff)
518 		index = 0;
519 
520 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
521 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
522 	addr += ((uint64_t)size) * ((uint64_t)index);
523 
524 	mapping = amdgpu_cs_find_mapping(p, addr, &bo);
525 	if (mapping == NULL) {
526 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
527 			  addr, lo, hi, size, index);
528 		return -EINVAL;
529 	}
530 
531 	if ((addr + (uint64_t)size) >
532 	    ((uint64_t)mapping->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
533 		DRM_ERROR("BO to small for addr 0x%010Lx %d %d\n",
534 			  addr, lo, hi);
535 		return -EINVAL;
536 	}
537 
538 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
539 	addr += amdgpu_bo_gpu_offset(bo);
540 	addr -= ((uint64_t)size) * ((uint64_t)index);
541 
542 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
543 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
544 
545 	return 0;
546 }
547 
548 /**
549  * amdgpu_vce_validate_handle - validate stream handle
550  *
551  * @p: parser context
552  * @handle: handle to validate
553  * @allocated: allocated a new handle?
554  *
555  * Validates the handle and return the found session index or -EINVAL
556  * we we don't have another free session index.
557  */
558 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
559 				      uint32_t handle, bool *allocated)
560 {
561 	unsigned i;
562 
563 	*allocated = false;
564 
565 	/* validate the handle */
566 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
567 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
568 			if (p->adev->vce.filp[i] != p->filp) {
569 				DRM_ERROR("VCE handle collision detected!\n");
570 				return -EINVAL;
571 			}
572 			return i;
573 		}
574 	}
575 
576 	/* handle not found try to alloc a new one */
577 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
578 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
579 			p->adev->vce.filp[i] = p->filp;
580 			p->adev->vce.img_size[i] = 0;
581 			*allocated = true;
582 			return i;
583 		}
584 	}
585 
586 	DRM_ERROR("No more free VCE handles!\n");
587 	return -EINVAL;
588 }
589 
590 /**
591  * amdgpu_vce_cs_parse - parse and validate the command stream
592  *
593  * @p: parser context
594  *
595  */
596 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
597 {
598 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
599 	unsigned fb_idx = 0, bs_idx = 0;
600 	int session_idx = -1;
601 	bool destroyed = false;
602 	bool created = false;
603 	bool allocated = false;
604 	uint32_t tmp, handle = 0;
605 	uint32_t *size = &tmp;
606 	int i, r = 0, idx = 0;
607 
608 	amdgpu_vce_note_usage(p->adev);
609 
610 	while (idx < ib->length_dw) {
611 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
612 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
613 
614 		if ((len < 8) || (len & 3)) {
615 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
616 			r = -EINVAL;
617 			goto out;
618 		}
619 
620 		if (destroyed) {
621 			DRM_ERROR("No other command allowed after destroy!\n");
622 			r = -EINVAL;
623 			goto out;
624 		}
625 
626 		switch (cmd) {
627 		case 0x00000001: // session
628 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
629 			session_idx = amdgpu_vce_validate_handle(p, handle,
630 								 &allocated);
631 			if (session_idx < 0)
632 				return session_idx;
633 			size = &p->adev->vce.img_size[session_idx];
634 			break;
635 
636 		case 0x00000002: // task info
637 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
638 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
639 			break;
640 
641 		case 0x01000001: // create
642 			created = true;
643 			if (!allocated) {
644 				DRM_ERROR("Handle already in use!\n");
645 				r = -EINVAL;
646 				goto out;
647 			}
648 
649 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
650 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
651 				8 * 3 / 2;
652 			break;
653 
654 		case 0x04000001: // config extension
655 		case 0x04000002: // pic control
656 		case 0x04000005: // rate control
657 		case 0x04000007: // motion estimation
658 		case 0x04000008: // rdo
659 		case 0x04000009: // vui
660 		case 0x05000002: // auxiliary buffer
661 			break;
662 
663 		case 0x03000001: // encode
664 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
665 						*size, 0);
666 			if (r)
667 				goto out;
668 
669 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
670 						*size / 3, 0);
671 			if (r)
672 				goto out;
673 			break;
674 
675 		case 0x02000001: // destroy
676 			destroyed = true;
677 			break;
678 
679 		case 0x05000001: // context buffer
680 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
681 						*size * 2, 0);
682 			if (r)
683 				goto out;
684 			break;
685 
686 		case 0x05000004: // video bitstream buffer
687 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
688 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
689 						tmp, bs_idx);
690 			if (r)
691 				goto out;
692 			break;
693 
694 		case 0x05000005: // feedback buffer
695 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
696 						4096, fb_idx);
697 			if (r)
698 				goto out;
699 			break;
700 
701 		default:
702 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
703 			r = -EINVAL;
704 			goto out;
705 		}
706 
707 		if (session_idx == -1) {
708 			DRM_ERROR("no session command at start of IB\n");
709 			r = -EINVAL;
710 			goto out;
711 		}
712 
713 		idx += len / 4;
714 	}
715 
716 	if (allocated && !created) {
717 		DRM_ERROR("New session without create command!\n");
718 		r = -ENOENT;
719 	}
720 
721 out:
722 	if ((!r && destroyed) || (r && allocated)) {
723 		/*
724 		 * IB contains a destroy msg or we have allocated an
725 		 * handle and got an error, anyway free the handle
726 		 */
727 		for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
728 			atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
729 	}
730 
731 	return r;
732 }
733 
734 /**
735  * amdgpu_vce_ring_emit_ib - execute indirect buffer
736  *
737  * @ring: engine to use
738  * @ib: the IB to execute
739  *
740  */
741 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
742 {
743 	amdgpu_ring_write(ring, VCE_CMD_IB);
744 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
745 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
746 	amdgpu_ring_write(ring, ib->length_dw);
747 }
748 
749 /**
750  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
751  *
752  * @ring: engine to use
753  * @fence: the fence
754  *
755  */
756 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
757 				unsigned flags)
758 {
759 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
760 
761 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
762 	amdgpu_ring_write(ring, addr);
763 	amdgpu_ring_write(ring, upper_32_bits(addr));
764 	amdgpu_ring_write(ring, seq);
765 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
766 	amdgpu_ring_write(ring, VCE_CMD_END);
767 }
768 
769 /**
770  * amdgpu_vce_ring_test_ring - test if VCE ring is working
771  *
772  * @ring: the engine to test on
773  *
774  */
775 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
776 {
777 	struct amdgpu_device *adev = ring->adev;
778 	uint32_t rptr = amdgpu_ring_get_rptr(ring);
779 	unsigned i;
780 	int r;
781 
782 	r = amdgpu_ring_alloc(ring, 16);
783 	if (r) {
784 		DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
785 			  ring->idx, r);
786 		return r;
787 	}
788 	amdgpu_ring_write(ring, VCE_CMD_END);
789 	amdgpu_ring_commit(ring);
790 
791 	for (i = 0; i < adev->usec_timeout; i++) {
792 		if (amdgpu_ring_get_rptr(ring) != rptr)
793 			break;
794 		DRM_UDELAY(1);
795 	}
796 
797 	if (i < adev->usec_timeout) {
798 		DRM_INFO("ring test on %d succeeded in %d usecs\n",
799 			 ring->idx, i);
800 	} else {
801 		DRM_ERROR("amdgpu: ring %d test failed\n",
802 			  ring->idx);
803 		r = -ETIMEDOUT;
804 	}
805 
806 	return r;
807 }
808 
809 /**
810  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
811  *
812  * @ring: the engine to test on
813  *
814  */
815 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
816 {
817 	struct fence *fence = NULL;
818 	int r;
819 
820 	/* skip vce ring1 ib test for now, since it's not reliable */
821 	if (ring == &ring->adev->vce.ring[1])
822 		return 0;
823 
824 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
825 	if (r) {
826 		DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
827 		goto error;
828 	}
829 
830 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
831 	if (r) {
832 		DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
833 		goto error;
834 	}
835 
836 	r = fence_wait(fence, false);
837 	if (r) {
838 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
839 	} else {
840 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
841 	}
842 error:
843 	fence_put(fence);
844 	return r;
845 }
846