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 	struct amdgpu_ring *ring;
78 	struct amd_sched_rq *rq;
79 	const char *fw_name;
80 	const struct common_firmware_header *hdr;
81 	unsigned ucode_version, version_major, version_minor, binary_id;
82 	int i, r;
83 
84 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
85 
86 	switch (adev->asic_type) {
87 #ifdef CONFIG_DRM_AMDGPU_CIK
88 	case CHIP_BONAIRE:
89 		fw_name = FIRMWARE_BONAIRE;
90 		break;
91 	case CHIP_KAVERI:
92 		fw_name = FIRMWARE_KAVERI;
93 		break;
94 	case CHIP_KABINI:
95 		fw_name = FIRMWARE_KABINI;
96 		break;
97 	case CHIP_HAWAII:
98 		fw_name = FIRMWARE_HAWAII;
99 		break;
100 	case CHIP_MULLINS:
101 		fw_name = FIRMWARE_MULLINS;
102 		break;
103 #endif
104 	case CHIP_TONGA:
105 		fw_name = FIRMWARE_TONGA;
106 		break;
107 	case CHIP_CARRIZO:
108 		fw_name = FIRMWARE_CARRIZO;
109 		break;
110 	case CHIP_FIJI:
111 		fw_name = FIRMWARE_FIJI;
112 		break;
113 	case CHIP_STONEY:
114 		fw_name = FIRMWARE_STONEY;
115 		break;
116 
117 	default:
118 		return -EINVAL;
119 	}
120 
121 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
122 	if (r) {
123 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
124 			fw_name);
125 		return r;
126 	}
127 
128 	r = amdgpu_ucode_validate(adev->vce.fw);
129 	if (r) {
130 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
131 			fw_name);
132 		release_firmware(adev->vce.fw);
133 		adev->vce.fw = NULL;
134 		return r;
135 	}
136 
137 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
138 
139 	ucode_version = le32_to_cpu(hdr->ucode_version);
140 	version_major = (ucode_version >> 20) & 0xfff;
141 	version_minor = (ucode_version >> 8) & 0xfff;
142 	binary_id = ucode_version & 0xff;
143 	DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
144 		version_major, version_minor, binary_id);
145 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
146 				(binary_id << 8));
147 
148 	/* allocate firmware, stack and heap BO */
149 
150 	r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
151 			     AMDGPU_GEM_DOMAIN_VRAM,
152 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
153 			     NULL, NULL, &adev->vce.vcpu_bo);
154 	if (r) {
155 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
156 		return r;
157 	}
158 
159 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
160 	if (r) {
161 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
162 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
163 		return r;
164 	}
165 
166 	r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
167 			  &adev->vce.gpu_addr);
168 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
169 	if (r) {
170 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
171 		dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
172 		return r;
173 	}
174 
175 
176 	ring = &adev->vce.ring[0];
177 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
178 	r = amd_sched_entity_init(&ring->sched, &adev->vce.entity,
179 				  rq, amdgpu_sched_jobs);
180 	if (r != 0) {
181 		DRM_ERROR("Failed setting up VCE run queue.\n");
182 		return r;
183 	}
184 
185 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
186 		atomic_set(&adev->vce.handles[i], 0);
187 		adev->vce.filp[i] = NULL;
188 	}
189 
190 	return 0;
191 }
192 
193 /**
194  * amdgpu_vce_fini - free memory
195  *
196  * @adev: amdgpu_device pointer
197  *
198  * Last step on VCE teardown, free firmware memory
199  */
200 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
201 {
202 	if (adev->vce.vcpu_bo == NULL)
203 		return 0;
204 
205 	amd_sched_entity_fini(&adev->vce.ring[0].sched, &adev->vce.entity);
206 
207 	amdgpu_bo_unref(&adev->vce.vcpu_bo);
208 
209 	amdgpu_ring_fini(&adev->vce.ring[0]);
210 	amdgpu_ring_fini(&adev->vce.ring[1]);
211 
212 	release_firmware(adev->vce.fw);
213 
214 	return 0;
215 }
216 
217 /**
218  * amdgpu_vce_suspend - unpin VCE fw memory
219  *
220  * @adev: amdgpu_device pointer
221  *
222  */
223 int amdgpu_vce_suspend(struct amdgpu_device *adev)
224 {
225 	int i;
226 
227 	if (adev->vce.vcpu_bo == NULL)
228 		return 0;
229 
230 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
231 		if (atomic_read(&adev->vce.handles[i]))
232 			break;
233 
234 	if (i == AMDGPU_MAX_VCE_HANDLES)
235 		return 0;
236 
237 	/* TODO: suspending running encoding sessions isn't supported */
238 	return -EINVAL;
239 }
240 
241 /**
242  * amdgpu_vce_resume - pin VCE fw memory
243  *
244  * @adev: amdgpu_device pointer
245  *
246  */
247 int amdgpu_vce_resume(struct amdgpu_device *adev)
248 {
249 	void *cpu_addr;
250 	const struct common_firmware_header *hdr;
251 	unsigned offset;
252 	int r;
253 
254 	if (adev->vce.vcpu_bo == NULL)
255 		return -EINVAL;
256 
257 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
258 	if (r) {
259 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
260 		return r;
261 	}
262 
263 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
264 	if (r) {
265 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
266 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
267 		return r;
268 	}
269 
270 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
271 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
272 	memcpy(cpu_addr, (adev->vce.fw->data) + offset,
273 		(adev->vce.fw->size) - offset);
274 
275 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
276 
277 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
278 
279 	return 0;
280 }
281 
282 /**
283  * amdgpu_vce_idle_work_handler - power off VCE
284  *
285  * @work: pointer to work structure
286  *
287  * power of VCE when it's not used any more
288  */
289 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
290 {
291 	struct amdgpu_device *adev =
292 		container_of(work, struct amdgpu_device, vce.idle_work.work);
293 
294 	if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
295 	    (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
296 		if (adev->pm.dpm_enabled) {
297 			amdgpu_dpm_enable_vce(adev, false);
298 		} else {
299 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
300 		}
301 	} else {
302 		schedule_delayed_work(&adev->vce.idle_work,
303 				      msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
304 	}
305 }
306 
307 /**
308  * amdgpu_vce_note_usage - power up VCE
309  *
310  * @adev: amdgpu_device pointer
311  *
312  * Make sure VCE is powerd up when we want to use it
313  */
314 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
315 {
316 	bool streams_changed = false;
317 	bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
318 	set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
319 					    msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
320 
321 	if (adev->pm.dpm_enabled) {
322 		/* XXX figure out if the streams changed */
323 		streams_changed = false;
324 	}
325 
326 	if (set_clocks || streams_changed) {
327 		if (adev->pm.dpm_enabled) {
328 			amdgpu_dpm_enable_vce(adev, true);
329 		} else {
330 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
331 		}
332 	}
333 }
334 
335 /**
336  * amdgpu_vce_free_handles - free still open VCE handles
337  *
338  * @adev: amdgpu_device pointer
339  * @filp: drm file pointer
340  *
341  * Close all VCE handles still open by this file pointer
342  */
343 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
344 {
345 	struct amdgpu_ring *ring = &adev->vce.ring[0];
346 	int i, r;
347 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
348 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
349 		if (!handle || adev->vce.filp[i] != filp)
350 			continue;
351 
352 		amdgpu_vce_note_usage(adev);
353 
354 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
355 		if (r)
356 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
357 
358 		adev->vce.filp[i] = NULL;
359 		atomic_set(&adev->vce.handles[i], 0);
360 	}
361 }
362 
363 /**
364  * amdgpu_vce_get_create_msg - generate a VCE create msg
365  *
366  * @adev: amdgpu_device pointer
367  * @ring: ring we should submit the msg to
368  * @handle: VCE session handle to use
369  * @fence: optional fence to return
370  *
371  * Open up a stream for HW test
372  */
373 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
374 			      struct fence **fence)
375 {
376 	const unsigned ib_size_dw = 1024;
377 	struct amdgpu_job *job;
378 	struct amdgpu_ib *ib;
379 	struct fence *f = NULL;
380 	uint64_t dummy;
381 	int i, r;
382 
383 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
384 	if (r)
385 		return r;
386 
387 	ib = &job->ibs[0];
388 
389 	dummy = ib->gpu_addr + 1024;
390 
391 	/* stitch together an VCE create msg */
392 	ib->length_dw = 0;
393 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
394 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
395 	ib->ptr[ib->length_dw++] = handle;
396 
397 	if ((ring->adev->vce.fw_version >> 24) >= 52)
398 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
399 	else
400 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
401 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
402 	ib->ptr[ib->length_dw++] = 0x00000000;
403 	ib->ptr[ib->length_dw++] = 0x00000042;
404 	ib->ptr[ib->length_dw++] = 0x0000000a;
405 	ib->ptr[ib->length_dw++] = 0x00000001;
406 	ib->ptr[ib->length_dw++] = 0x00000080;
407 	ib->ptr[ib->length_dw++] = 0x00000060;
408 	ib->ptr[ib->length_dw++] = 0x00000100;
409 	ib->ptr[ib->length_dw++] = 0x00000100;
410 	ib->ptr[ib->length_dw++] = 0x0000000c;
411 	ib->ptr[ib->length_dw++] = 0x00000000;
412 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
413 		ib->ptr[ib->length_dw++] = 0x00000000;
414 		ib->ptr[ib->length_dw++] = 0x00000000;
415 		ib->ptr[ib->length_dw++] = 0x00000000;
416 		ib->ptr[ib->length_dw++] = 0x00000000;
417 	}
418 
419 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
420 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
421 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
422 	ib->ptr[ib->length_dw++] = dummy;
423 	ib->ptr[ib->length_dw++] = 0x00000001;
424 
425 	for (i = ib->length_dw; i < ib_size_dw; ++i)
426 		ib->ptr[i] = 0x0;
427 
428 	r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
429 	job->fence = f;
430 	if (r)
431 		goto err;
432 
433 	amdgpu_job_free(job);
434 	if (fence)
435 		*fence = fence_get(f);
436 	fence_put(f);
437 	return 0;
438 
439 err:
440 	amdgpu_job_free(job);
441 	return r;
442 }
443 
444 /**
445  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
446  *
447  * @adev: amdgpu_device pointer
448  * @ring: ring we should submit the msg to
449  * @handle: VCE session handle to use
450  * @fence: optional fence to return
451  *
452  * Close up a stream for HW test or if userspace failed to do so
453  */
454 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
455 			       bool direct, struct fence **fence)
456 {
457 	const unsigned ib_size_dw = 1024;
458 	struct amdgpu_job *job;
459 	struct amdgpu_ib *ib;
460 	struct fence *f = NULL;
461 	uint64_t dummy;
462 	int i, r;
463 
464 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
465 	if (r)
466 		return r;
467 
468 	ib = &job->ibs[0];
469 	dummy = ib->gpu_addr + 1024;
470 
471 	/* stitch together an VCE destroy msg */
472 	ib->length_dw = 0;
473 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
474 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
475 	ib->ptr[ib->length_dw++] = handle;
476 
477 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
478 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
479 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
480 	ib->ptr[ib->length_dw++] = dummy;
481 	ib->ptr[ib->length_dw++] = 0x00000001;
482 
483 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
484 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
485 
486 	for (i = ib->length_dw; i < ib_size_dw; ++i)
487 		ib->ptr[i] = 0x0;
488 
489 	if (direct) {
490 		r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
491 		job->fence = f;
492 		if (r)
493 			goto err;
494 
495 		amdgpu_job_free(job);
496 	} else {
497 		r = amdgpu_job_submit(job, ring, &ring->adev->vce.entity,
498 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
499 		if (r)
500 			goto err;
501 	}
502 
503 	if (fence)
504 		*fence = fence_get(f);
505 	fence_put(f);
506 	return 0;
507 
508 err:
509 	amdgpu_job_free(job);
510 	return r;
511 }
512 
513 /**
514  * amdgpu_vce_cs_reloc - command submission relocation
515  *
516  * @p: parser context
517  * @lo: address of lower dword
518  * @hi: address of higher dword
519  * @size: minimum size
520  *
521  * Patch relocation inside command stream with real buffer address
522  */
523 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
524 			       int lo, int hi, unsigned size, uint32_t index)
525 {
526 	struct amdgpu_bo_va_mapping *mapping;
527 	struct amdgpu_bo *bo;
528 	uint64_t addr;
529 
530 	if (index == 0xffffffff)
531 		index = 0;
532 
533 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
534 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
535 	addr += ((uint64_t)size) * ((uint64_t)index);
536 
537 	mapping = amdgpu_cs_find_mapping(p, addr, &bo);
538 	if (mapping == NULL) {
539 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
540 			  addr, lo, hi, size, index);
541 		return -EINVAL;
542 	}
543 
544 	if ((addr + (uint64_t)size) >
545 	    ((uint64_t)mapping->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
546 		DRM_ERROR("BO to small for addr 0x%010Lx %d %d\n",
547 			  addr, lo, hi);
548 		return -EINVAL;
549 	}
550 
551 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
552 	addr += amdgpu_bo_gpu_offset(bo);
553 	addr -= ((uint64_t)size) * ((uint64_t)index);
554 
555 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
556 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
557 
558 	return 0;
559 }
560 
561 /**
562  * amdgpu_vce_validate_handle - validate stream handle
563  *
564  * @p: parser context
565  * @handle: handle to validate
566  * @allocated: allocated a new handle?
567  *
568  * Validates the handle and return the found session index or -EINVAL
569  * we we don't have another free session index.
570  */
571 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
572 				      uint32_t handle, bool *allocated)
573 {
574 	unsigned i;
575 
576 	*allocated = false;
577 
578 	/* validate the handle */
579 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
580 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
581 			if (p->adev->vce.filp[i] != p->filp) {
582 				DRM_ERROR("VCE handle collision detected!\n");
583 				return -EINVAL;
584 			}
585 			return i;
586 		}
587 	}
588 
589 	/* handle not found try to alloc a new one */
590 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
591 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
592 			p->adev->vce.filp[i] = p->filp;
593 			p->adev->vce.img_size[i] = 0;
594 			*allocated = true;
595 			return i;
596 		}
597 	}
598 
599 	DRM_ERROR("No more free VCE handles!\n");
600 	return -EINVAL;
601 }
602 
603 /**
604  * amdgpu_vce_cs_parse - parse and validate the command stream
605  *
606  * @p: parser context
607  *
608  */
609 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
610 {
611 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
612 	unsigned fb_idx = 0, bs_idx = 0;
613 	int session_idx = -1;
614 	bool destroyed = false;
615 	bool created = false;
616 	bool allocated = false;
617 	uint32_t tmp, handle = 0;
618 	uint32_t *size = &tmp;
619 	int i, r = 0, idx = 0;
620 
621 	amdgpu_vce_note_usage(p->adev);
622 
623 	while (idx < ib->length_dw) {
624 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
625 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
626 
627 		if ((len < 8) || (len & 3)) {
628 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
629 			r = -EINVAL;
630 			goto out;
631 		}
632 
633 		if (destroyed) {
634 			DRM_ERROR("No other command allowed after destroy!\n");
635 			r = -EINVAL;
636 			goto out;
637 		}
638 
639 		switch (cmd) {
640 		case 0x00000001: // session
641 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
642 			session_idx = amdgpu_vce_validate_handle(p, handle,
643 								 &allocated);
644 			if (session_idx < 0)
645 				return session_idx;
646 			size = &p->adev->vce.img_size[session_idx];
647 			break;
648 
649 		case 0x00000002: // task info
650 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
651 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
652 			break;
653 
654 		case 0x01000001: // create
655 			created = true;
656 			if (!allocated) {
657 				DRM_ERROR("Handle already in use!\n");
658 				r = -EINVAL;
659 				goto out;
660 			}
661 
662 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
663 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
664 				8 * 3 / 2;
665 			break;
666 
667 		case 0x04000001: // config extension
668 		case 0x04000002: // pic control
669 		case 0x04000005: // rate control
670 		case 0x04000007: // motion estimation
671 		case 0x04000008: // rdo
672 		case 0x04000009: // vui
673 		case 0x05000002: // auxiliary buffer
674 			break;
675 
676 		case 0x03000001: // encode
677 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
678 						*size, 0);
679 			if (r)
680 				goto out;
681 
682 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
683 						*size / 3, 0);
684 			if (r)
685 				goto out;
686 			break;
687 
688 		case 0x02000001: // destroy
689 			destroyed = true;
690 			break;
691 
692 		case 0x05000001: // context buffer
693 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
694 						*size * 2, 0);
695 			if (r)
696 				goto out;
697 			break;
698 
699 		case 0x05000004: // video bitstream buffer
700 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
701 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
702 						tmp, bs_idx);
703 			if (r)
704 				goto out;
705 			break;
706 
707 		case 0x05000005: // feedback buffer
708 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
709 						4096, fb_idx);
710 			if (r)
711 				goto out;
712 			break;
713 
714 		default:
715 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
716 			r = -EINVAL;
717 			goto out;
718 		}
719 
720 		if (session_idx == -1) {
721 			DRM_ERROR("no session command at start of IB\n");
722 			r = -EINVAL;
723 			goto out;
724 		}
725 
726 		idx += len / 4;
727 	}
728 
729 	if (allocated && !created) {
730 		DRM_ERROR("New session without create command!\n");
731 		r = -ENOENT;
732 	}
733 
734 out:
735 	if ((!r && destroyed) || (r && allocated)) {
736 		/*
737 		 * IB contains a destroy msg or we have allocated an
738 		 * handle and got an error, anyway free the handle
739 		 */
740 		for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
741 			atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
742 	}
743 
744 	return r;
745 }
746 
747 /**
748  * amdgpu_vce_ring_emit_ib - execute indirect buffer
749  *
750  * @ring: engine to use
751  * @ib: the IB to execute
752  *
753  */
754 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
755 {
756 	amdgpu_ring_write(ring, VCE_CMD_IB);
757 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
758 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
759 	amdgpu_ring_write(ring, ib->length_dw);
760 }
761 
762 /**
763  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
764  *
765  * @ring: engine to use
766  * @fence: the fence
767  *
768  */
769 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
770 				unsigned flags)
771 {
772 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
773 
774 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
775 	amdgpu_ring_write(ring, addr);
776 	amdgpu_ring_write(ring, upper_32_bits(addr));
777 	amdgpu_ring_write(ring, seq);
778 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
779 	amdgpu_ring_write(ring, VCE_CMD_END);
780 }
781 
782 /**
783  * amdgpu_vce_ring_test_ring - test if VCE ring is working
784  *
785  * @ring: the engine to test on
786  *
787  */
788 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
789 {
790 	struct amdgpu_device *adev = ring->adev;
791 	uint32_t rptr = amdgpu_ring_get_rptr(ring);
792 	unsigned i;
793 	int r;
794 
795 	r = amdgpu_ring_alloc(ring, 16);
796 	if (r) {
797 		DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
798 			  ring->idx, r);
799 		return r;
800 	}
801 	amdgpu_ring_write(ring, VCE_CMD_END);
802 	amdgpu_ring_commit(ring);
803 
804 	for (i = 0; i < adev->usec_timeout; i++) {
805 		if (amdgpu_ring_get_rptr(ring) != rptr)
806 			break;
807 		DRM_UDELAY(1);
808 	}
809 
810 	if (i < adev->usec_timeout) {
811 		DRM_INFO("ring test on %d succeeded in %d usecs\n",
812 			 ring->idx, i);
813 	} else {
814 		DRM_ERROR("amdgpu: ring %d test failed\n",
815 			  ring->idx);
816 		r = -ETIMEDOUT;
817 	}
818 
819 	return r;
820 }
821 
822 /**
823  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
824  *
825  * @ring: the engine to test on
826  *
827  */
828 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
829 {
830 	struct fence *fence = NULL;
831 	int r;
832 
833 	/* skip vce ring1 ib test for now, since it's not reliable */
834 	if (ring == &ring->adev->vce.ring[1])
835 		return 0;
836 
837 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
838 	if (r) {
839 		DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
840 		goto error;
841 	}
842 
843 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
844 	if (r) {
845 		DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
846 		goto error;
847 	}
848 
849 	r = fence_wait(fence, false);
850 	if (r) {
851 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
852 	} else {
853 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
854 	}
855 error:
856 	fence_put(fence);
857 	return r;
858 }
859