xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_vce.c (revision f3cfa6f6ce31685c6c4a758bc430e69eb99f50a4)
1 /*	$NetBSD: amdgpu_vce.c,v 1.4 2019/01/01 08:07:47 maya Exp $	*/
2 
3 /*
4  * Copyright 2013 Advanced Micro Devices, Inc.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * The above copyright notice and this permission notice (including the
24  * next paragraph) shall be included in all copies or substantial portions
25  * of the Software.
26  *
27  * Authors: Christian König <christian.koenig@amd.com>
28  */
29 
30 #include <sys/cdefs.h>
31 __KERNEL_RCSID(0, "$NetBSD: amdgpu_vce.c,v 1.4 2019/01/01 08:07:47 maya Exp $");
32 
33 #include <linux/firmware.h>
34 #include <linux/module.h>
35 #include <asm/byteorder.h>
36 #include <drm/drmP.h>
37 #include <drm/drm.h>
38 
39 #include "amdgpu.h"
40 #include "amdgpu_pm.h"
41 #include "amdgpu_vce.h"
42 #include "cikd.h"
43 
44 /* 1 second timeout */
45 #define VCE_IDLE_TIMEOUT_MS	1000
46 
47 /* Firmware Names */
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 #define FIRMWARE_BONAIRE	"radeon/bonaire_vce.bin"
50 #define FIRMWARE_KABINI 	"radeon/kabini_vce.bin"
51 #define FIRMWARE_KAVERI 	"radeon/kaveri_vce.bin"
52 #define FIRMWARE_HAWAII 	"radeon/hawaii_vce.bin"
53 #define FIRMWARE_MULLINS	"radeon/mullins_vce.bin"
54 #endif
55 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
56 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
57 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
58 #define FIRMWARE_STONEY		"amdgpu/stoney_vce.bin"
59 
60 #ifdef CONFIG_DRM_AMDGPU_CIK
61 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
62 MODULE_FIRMWARE(FIRMWARE_KABINI);
63 MODULE_FIRMWARE(FIRMWARE_KAVERI);
64 MODULE_FIRMWARE(FIRMWARE_HAWAII);
65 MODULE_FIRMWARE(FIRMWARE_MULLINS);
66 #endif
67 MODULE_FIRMWARE(FIRMWARE_TONGA);
68 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
69 MODULE_FIRMWARE(FIRMWARE_FIJI);
70 MODULE_FIRMWARE(FIRMWARE_STONEY);
71 
72 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
73 
74 /**
75  * amdgpu_vce_init - allocate memory, load vce firmware
76  *
77  * @adev: amdgpu_device pointer
78  *
79  * First step to get VCE online, allocate memory and load the firmware
80  */
81 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
82 {
83 	const char *fw_name;
84 	const struct common_firmware_header *hdr;
85 	unsigned ucode_version, version_major, version_minor, binary_id;
86 	int i, r;
87 
88 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
89 
90 	switch (adev->asic_type) {
91 #ifdef CONFIG_DRM_AMDGPU_CIK
92 	case CHIP_BONAIRE:
93 		fw_name = FIRMWARE_BONAIRE;
94 		break;
95 	case CHIP_KAVERI:
96 		fw_name = FIRMWARE_KAVERI;
97 		break;
98 	case CHIP_KABINI:
99 		fw_name = FIRMWARE_KABINI;
100 		break;
101 	case CHIP_HAWAII:
102 		fw_name = FIRMWARE_HAWAII;
103 		break;
104 	case CHIP_MULLINS:
105 		fw_name = FIRMWARE_MULLINS;
106 		break;
107 #endif
108 	case CHIP_TONGA:
109 		fw_name = FIRMWARE_TONGA;
110 		break;
111 	case CHIP_CARRIZO:
112 		fw_name = FIRMWARE_CARRIZO;
113 		break;
114 	case CHIP_FIJI:
115 		fw_name = FIRMWARE_FIJI;
116 		break;
117 	case CHIP_STONEY:
118 		fw_name = FIRMWARE_STONEY;
119 		break;
120 
121 	default:
122 		return -EINVAL;
123 	}
124 
125 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
126 	if (r) {
127 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
128 			fw_name);
129 		return r;
130 	}
131 
132 	r = amdgpu_ucode_validate(adev->vce.fw);
133 	if (r) {
134 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
135 			fw_name);
136 		release_firmware(adev->vce.fw);
137 		adev->vce.fw = NULL;
138 		return r;
139 	}
140 
141 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
142 
143 	ucode_version = le32_to_cpu(hdr->ucode_version);
144 	version_major = (ucode_version >> 20) & 0xfff;
145 	version_minor = (ucode_version >> 8) & 0xfff;
146 	binary_id = ucode_version & 0xff;
147 	DRM_INFO("Found VCE firmware Version: %x.%x Binary ID: %x\n",
148 		version_major, version_minor, binary_id);
149 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
150 				(binary_id << 8));
151 
152 	/* allocate firmware, stack and heap BO */
153 
154 	r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
155 			     AMDGPU_GEM_DOMAIN_VRAM,
156 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
157 			     NULL, NULL, &adev->vce.vcpu_bo);
158 	if (r) {
159 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
160 		return r;
161 	}
162 
163 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
164 	if (r) {
165 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
166 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
167 		return r;
168 	}
169 
170 	r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
171 			  &adev->vce.gpu_addr);
172 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
173 	if (r) {
174 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
175 		dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
176 		return r;
177 	}
178 
179 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
180 		atomic_set(&adev->vce.handles[i], 0);
181 		adev->vce.filp[i] = NULL;
182 	}
183 
184 	return 0;
185 }
186 
187 /**
188  * amdgpu_vce_fini - free memory
189  *
190  * @adev: amdgpu_device pointer
191  *
192  * Last step on VCE teardown, free firmware memory
193  */
194 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
195 {
196 	if (adev->vce.vcpu_bo == NULL)
197 		return 0;
198 
199 	amdgpu_bo_unref(&adev->vce.vcpu_bo);
200 
201 	amdgpu_ring_fini(&adev->vce.ring[0]);
202 	amdgpu_ring_fini(&adev->vce.ring[1]);
203 
204 	release_firmware(adev->vce.fw);
205 
206 	return 0;
207 }
208 
209 /**
210  * amdgpu_vce_suspend - unpin VCE fw memory
211  *
212  * @adev: amdgpu_device pointer
213  *
214  */
215 int amdgpu_vce_suspend(struct amdgpu_device *adev)
216 {
217 	int i;
218 
219 	if (adev->vce.vcpu_bo == NULL)
220 		return 0;
221 
222 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
223 		if (atomic_read(&adev->vce.handles[i]))
224 			break;
225 
226 	if (i == AMDGPU_MAX_VCE_HANDLES)
227 		return 0;
228 
229 	cancel_delayed_work_sync(&adev->vce.idle_work);
230 	/* TODO: suspending running encoding sessions isn't supported */
231 	return -EINVAL;
232 }
233 
234 /**
235  * amdgpu_vce_resume - pin VCE fw memory
236  *
237  * @adev: amdgpu_device pointer
238  *
239  */
240 int amdgpu_vce_resume(struct amdgpu_device *adev)
241 {
242 	void *cpu_addr;
243 	const struct common_firmware_header *hdr;
244 	unsigned offset;
245 	int r;
246 
247 	if (adev->vce.vcpu_bo == NULL)
248 		return -EINVAL;
249 
250 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
251 	if (r) {
252 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
253 		return r;
254 	}
255 
256 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
257 	if (r) {
258 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
259 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
260 		return r;
261 	}
262 
263 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
264 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
265 	memcpy(cpu_addr, (adev->vce.fw->data) + offset,
266 		(adev->vce.fw->size) - offset);
267 
268 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
269 
270 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
271 
272 	return 0;
273 }
274 
275 /**
276  * amdgpu_vce_idle_work_handler - power off VCE
277  *
278  * @work: pointer to work structure
279  *
280  * power of VCE when it's not used any more
281  */
282 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
283 {
284 	struct amdgpu_device *adev =
285 		container_of(work, struct amdgpu_device, vce.idle_work.work);
286 
287 	if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
288 	    (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
289 		if (adev->pm.dpm_enabled) {
290 			amdgpu_dpm_enable_vce(adev, false);
291 		} else {
292 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
293 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
294 							    AMD_PG_STATE_GATE);
295 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
296 							    AMD_CG_STATE_GATE);
297 		}
298 	} else {
299 		schedule_delayed_work(&adev->vce.idle_work,
300 				      msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
301 	}
302 }
303 
304 /**
305  * amdgpu_vce_note_usage - power up VCE
306  *
307  * @adev: amdgpu_device pointer
308  *
309  * Make sure VCE is powerd up when we want to use it
310  */
311 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
312 {
313 	bool streams_changed = false;
314 	bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
315 	set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
316 					    msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
317 
318 	if (adev->pm.dpm_enabled) {
319 		/* XXX figure out if the streams changed */
320 		streams_changed = false;
321 	}
322 
323 	if (set_clocks || streams_changed) {
324 		if (adev->pm.dpm_enabled) {
325 			amdgpu_dpm_enable_vce(adev, true);
326 		} else {
327 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
328 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
329 							    AMD_CG_STATE_UNGATE);
330 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
331 							    AMD_PG_STATE_UNGATE);
332 
333 		}
334 	}
335 }
336 
337 /**
338  * amdgpu_vce_free_handles - free still open VCE handles
339  *
340  * @adev: amdgpu_device pointer
341  * @filp: drm file pointer
342  *
343  * Close all VCE handles still open by this file pointer
344  */
345 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
346 {
347 	struct amdgpu_ring *ring = &adev->vce.ring[0];
348 	int i, r;
349 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
350 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
351 		if (!handle || adev->vce.filp[i] != filp)
352 			continue;
353 
354 		amdgpu_vce_note_usage(adev);
355 
356 		r = amdgpu_vce_get_destroy_msg(ring, handle, NULL);
357 		if (r)
358 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
359 
360 		adev->vce.filp[i] = NULL;
361 		atomic_set(&adev->vce.handles[i], 0);
362 	}
363 }
364 
365 static int amdgpu_vce_free_job(
366 	struct amdgpu_job *job)
367 {
368 	amdgpu_ib_free(job->adev, job->ibs);
369 	kfree(job->ibs);
370 	return 0;
371 }
372 
373 /**
374  * amdgpu_vce_get_create_msg - generate a VCE create msg
375  *
376  * @adev: amdgpu_device pointer
377  * @ring: ring we should submit the msg to
378  * @handle: VCE session handle to use
379  * @fence: optional fence to return
380  *
381  * Open up a stream for HW test
382  */
383 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
384 			      struct fence **fence)
385 {
386 	const unsigned ib_size_dw = 1024;
387 	struct amdgpu_ib *ib = NULL;
388 	struct fence *f = NULL;
389 	struct amdgpu_device *adev = ring->adev;
390 	uint64_t dummy;
391 	int i, r;
392 
393 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
394 	if (!ib)
395 		return -ENOMEM;
396 	r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
397 	if (r) {
398 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
399 		kfree(ib);
400 		return r;
401 	}
402 
403 	dummy = ib->gpu_addr + 1024;
404 
405 	/* stitch together an VCE create msg */
406 	ib->length_dw = 0;
407 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
408 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
409 	ib->ptr[ib->length_dw++] = handle;
410 
411 	if ((ring->adev->vce.fw_version >> 24) >= 52)
412 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
413 	else
414 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
415 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
416 	ib->ptr[ib->length_dw++] = 0x00000000;
417 	ib->ptr[ib->length_dw++] = 0x00000042;
418 	ib->ptr[ib->length_dw++] = 0x0000000a;
419 	ib->ptr[ib->length_dw++] = 0x00000001;
420 	ib->ptr[ib->length_dw++] = 0x00000080;
421 	ib->ptr[ib->length_dw++] = 0x00000060;
422 	ib->ptr[ib->length_dw++] = 0x00000100;
423 	ib->ptr[ib->length_dw++] = 0x00000100;
424 	ib->ptr[ib->length_dw++] = 0x0000000c;
425 	ib->ptr[ib->length_dw++] = 0x00000000;
426 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
427 		ib->ptr[ib->length_dw++] = 0x00000000;
428 		ib->ptr[ib->length_dw++] = 0x00000000;
429 		ib->ptr[ib->length_dw++] = 0x00000000;
430 		ib->ptr[ib->length_dw++] = 0x00000000;
431 	}
432 
433 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
434 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
435 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
436 	ib->ptr[ib->length_dw++] = dummy;
437 	ib->ptr[ib->length_dw++] = 0x00000001;
438 
439 	for (i = ib->length_dw; i < ib_size_dw; ++i)
440 		ib->ptr[i] = 0x0;
441 
442 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
443 						 &amdgpu_vce_free_job,
444 						 AMDGPU_FENCE_OWNER_UNDEFINED,
445 						 &f);
446 	if (r)
447 		goto err;
448 	if (fence)
449 		*fence = fence_get(f);
450 	fence_put(f);
451 	if (amdgpu_enable_scheduler)
452 		return 0;
453 err:
454 	amdgpu_ib_free(adev, ib);
455 	kfree(ib);
456 	return r;
457 }
458 
459 /**
460  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
461  *
462  * @adev: amdgpu_device pointer
463  * @ring: ring we should submit the msg to
464  * @handle: VCE session handle to use
465  * @fence: optional fence to return
466  *
467  * Close up a stream for HW test or if userspace failed to do so
468  */
469 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
470 			       struct fence **fence)
471 {
472 	const unsigned ib_size_dw = 1024;
473 	struct amdgpu_ib *ib = NULL;
474 	struct fence *f = NULL;
475 	struct amdgpu_device *adev = ring->adev;
476 	uint64_t dummy;
477 	int i, r;
478 
479 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
480 	if (!ib)
481 		return -ENOMEM;
482 
483 	r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
484 	if (r) {
485 		kfree(ib);
486 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
487 		return r;
488 	}
489 
490 	dummy = ib->gpu_addr + 1024;
491 
492 	/* stitch together an VCE destroy msg */
493 	ib->length_dw = 0;
494 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
495 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
496 	ib->ptr[ib->length_dw++] = handle;
497 
498 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
499 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
500 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
501 	ib->ptr[ib->length_dw++] = dummy;
502 	ib->ptr[ib->length_dw++] = 0x00000001;
503 
504 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
505 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
506 
507 	for (i = ib->length_dw; i < ib_size_dw; ++i)
508 		ib->ptr[i] = 0x0;
509 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
510 						 &amdgpu_vce_free_job,
511 						 AMDGPU_FENCE_OWNER_UNDEFINED,
512 						 &f);
513 	if (r)
514 		goto err;
515 	if (fence)
516 		*fence = fence_get(f);
517 	fence_put(f);
518 	if (amdgpu_enable_scheduler)
519 		return 0;
520 err:
521 	amdgpu_ib_free(adev, ib);
522 	kfree(ib);
523 	return r;
524 }
525 
526 /**
527  * amdgpu_vce_cs_reloc - command submission relocation
528  *
529  * @p: parser context
530  * @lo: address of lower dword
531  * @hi: address of higher dword
532  * @size: minimum size
533  *
534  * Patch relocation inside command stream with real buffer address
535  */
536 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
537 			       int lo, int hi, unsigned size, uint32_t index)
538 {
539 	struct amdgpu_bo_va_mapping *mapping;
540 	struct amdgpu_ib *ib = &p->ibs[ib_idx];
541 	struct amdgpu_bo *bo;
542 	uint64_t addr;
543 
544 	if (index == 0xffffffff)
545 		index = 0;
546 
547 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
548 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
549 	addr += ((uint64_t)size) * ((uint64_t)index);
550 
551 	mapping = amdgpu_cs_find_mapping(p, addr, &bo);
552 	if (mapping == NULL) {
553 		DRM_ERROR("Can't find BO for addr 0x%010"PRIx64" %d %d %d %d\n",
554 			  addr, lo, hi, size, index);
555 		return -EINVAL;
556 	}
557 
558 	if ((addr + (uint64_t)size) >
559 	    ((uint64_t)mapping->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
560 		DRM_ERROR("BO to small for addr 0x%010"PRIx64" %d %d\n",
561 			  addr, lo, hi);
562 		return -EINVAL;
563 	}
564 
565 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
566 	addr += amdgpu_bo_gpu_offset(bo);
567 	addr -= ((uint64_t)size) * ((uint64_t)index);
568 
569 	ib->ptr[lo] = addr & 0xFFFFFFFF;
570 	ib->ptr[hi] = addr >> 32;
571 
572 	return 0;
573 }
574 
575 /**
576  * amdgpu_vce_validate_handle - validate stream handle
577  *
578  * @p: parser context
579  * @handle: handle to validate
580  * @allocated: allocated a new handle?
581  *
582  * Validates the handle and return the found session index or -EINVAL
583  * we we don't have another free session index.
584  */
585 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
586 				      uint32_t handle, bool *allocated)
587 {
588 	unsigned i;
589 
590 	*allocated = false;
591 
592 	/* validate the handle */
593 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
594 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
595 			if (p->adev->vce.filp[i] != p->filp) {
596 				DRM_ERROR("VCE handle collision detected!\n");
597 				return -EINVAL;
598 			}
599 			return i;
600 		}
601 	}
602 
603 	/* handle not found try to alloc a new one */
604 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
605 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
606 			p->adev->vce.filp[i] = p->filp;
607 			p->adev->vce.img_size[i] = 0;
608 			*allocated = true;
609 			return i;
610 		}
611 	}
612 
613 	DRM_ERROR("No more free VCE handles!\n");
614 	return -EINVAL;
615 }
616 
617 /**
618  * amdgpu_vce_cs_parse - parse and validate the command stream
619  *
620  * @p: parser context
621  *
622  */
623 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
624 {
625 	struct amdgpu_ib *ib = &p->ibs[ib_idx];
626 	unsigned fb_idx = 0, bs_idx = 0;
627 	int session_idx = -1;
628 	bool destroyed = false;
629 	bool created = false;
630 	bool allocated = false;
631 	uint32_t tmp, handle = 0;
632 	uint32_t *size = &tmp;
633 	int i, r = 0, idx = 0;
634 
635 	amdgpu_vce_note_usage(p->adev);
636 
637 	while (idx < ib->length_dw) {
638 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
639 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
640 
641 		if ((len < 8) || (len & 3)) {
642 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
643 			r = -EINVAL;
644 			goto out;
645 		}
646 
647 		if (destroyed) {
648 			DRM_ERROR("No other command allowed after destroy!\n");
649 			r = -EINVAL;
650 			goto out;
651 		}
652 
653 		switch (cmd) {
654 		case 0x00000001: // session
655 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
656 			session_idx = amdgpu_vce_validate_handle(p, handle,
657 								 &allocated);
658 			if (session_idx < 0)
659 				return session_idx;
660 			size = &p->adev->vce.img_size[session_idx];
661 			break;
662 
663 		case 0x00000002: // task info
664 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
665 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
666 			break;
667 
668 		case 0x01000001: // create
669 			created = true;
670 			if (!allocated) {
671 				DRM_ERROR("Handle already in use!\n");
672 				r = -EINVAL;
673 				goto out;
674 			}
675 
676 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
677 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
678 				8 * 3 / 2;
679 			break;
680 
681 		case 0x04000001: // config extension
682 		case 0x04000002: // pic control
683 		case 0x04000005: // rate control
684 		case 0x04000007: // motion estimation
685 		case 0x04000008: // rdo
686 		case 0x04000009: // vui
687 		case 0x05000002: // auxiliary buffer
688 			break;
689 
690 		case 0x03000001: // encode
691 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
692 						*size, 0);
693 			if (r)
694 				goto out;
695 
696 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
697 						*size / 3, 0);
698 			if (r)
699 				goto out;
700 			break;
701 
702 		case 0x02000001: // destroy
703 			destroyed = true;
704 			break;
705 
706 		case 0x05000001: // context buffer
707 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
708 						*size * 2, 0);
709 			if (r)
710 				goto out;
711 			break;
712 
713 		case 0x05000004: // video bitstream buffer
714 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
715 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
716 						tmp, bs_idx);
717 			if (r)
718 				goto out;
719 			break;
720 
721 		case 0x05000005: // feedback buffer
722 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
723 						4096, fb_idx);
724 			if (r)
725 				goto out;
726 			break;
727 
728 		default:
729 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
730 			r = -EINVAL;
731 			goto out;
732 		}
733 
734 		if (session_idx == -1) {
735 			DRM_ERROR("no session command at start of IB\n");
736 			r = -EINVAL;
737 			goto out;
738 		}
739 
740 		idx += len / 4;
741 	}
742 
743 	if (allocated && !created) {
744 		DRM_ERROR("New session without create command!\n");
745 		r = -ENOENT;
746 	}
747 
748 out:
749 	if ((!r && destroyed) || (r && allocated)) {
750 		/*
751 		 * IB contains a destroy msg or we have allocated an
752 		 * handle and got an error, anyway free the handle
753 		 */
754 		for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
755 			atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
756 	}
757 
758 	return r;
759 }
760 
761 /**
762  * amdgpu_vce_ring_emit_semaphore - emit a semaphore command
763  *
764  * @ring: engine to use
765  * @semaphore: address of semaphore
766  * @emit_wait: true=emit wait, false=emit signal
767  *
768  */
769 bool amdgpu_vce_ring_emit_semaphore(struct amdgpu_ring *ring,
770 				    struct amdgpu_semaphore *semaphore,
771 				    bool emit_wait)
772 {
773 	uint64_t addr = semaphore->gpu_addr;
774 
775 	amdgpu_ring_write(ring, VCE_CMD_SEMAPHORE);
776 	amdgpu_ring_write(ring, (addr >> 3) & 0x000FFFFF);
777 	amdgpu_ring_write(ring, (addr >> 23) & 0x000FFFFF);
778 	amdgpu_ring_write(ring, 0x01003000 | (emit_wait ? 1 : 0));
779 	if (!emit_wait)
780 		amdgpu_ring_write(ring, VCE_CMD_END);
781 
782 	return true;
783 }
784 
785 /**
786  * amdgpu_vce_ring_emit_ib - execute indirect buffer
787  *
788  * @ring: engine to use
789  * @ib: the IB to execute
790  *
791  */
792 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
793 {
794 	amdgpu_ring_write(ring, VCE_CMD_IB);
795 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
796 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
797 	amdgpu_ring_write(ring, ib->length_dw);
798 }
799 
800 /**
801  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
802  *
803  * @ring: engine to use
804  * @fence: the fence
805  *
806  */
807 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
808 				unsigned flags)
809 {
810 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
811 
812 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
813 	amdgpu_ring_write(ring, addr);
814 	amdgpu_ring_write(ring, upper_32_bits(addr));
815 	amdgpu_ring_write(ring, seq);
816 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
817 	amdgpu_ring_write(ring, VCE_CMD_END);
818 }
819 
820 /**
821  * amdgpu_vce_ring_test_ring - test if VCE ring is working
822  *
823  * @ring: the engine to test on
824  *
825  */
826 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
827 {
828 	struct amdgpu_device *adev = ring->adev;
829 	uint32_t rptr = amdgpu_ring_get_rptr(ring);
830 	unsigned i;
831 	int r;
832 
833 	r = amdgpu_ring_lock(ring, 16);
834 	if (r) {
835 		DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
836 			  ring->idx, r);
837 		return r;
838 	}
839 	amdgpu_ring_write(ring, VCE_CMD_END);
840 	amdgpu_ring_unlock_commit(ring);
841 
842 	for (i = 0; i < adev->usec_timeout; i++) {
843 		if (amdgpu_ring_get_rptr(ring) != rptr)
844 			break;
845 		DRM_UDELAY(1);
846 	}
847 
848 	if (i < adev->usec_timeout) {
849 		DRM_INFO("ring test on %d succeeded in %d usecs\n",
850 			 ring->idx, i);
851 	} else {
852 		DRM_ERROR("amdgpu: ring %d test failed\n",
853 			  ring->idx);
854 		r = -ETIMEDOUT;
855 	}
856 
857 	return r;
858 }
859 
860 /**
861  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
862  *
863  * @ring: the engine to test on
864  *
865  */
866 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
867 {
868 	struct fence *fence = NULL;
869 	int r;
870 
871 	/* skip vce ring1 ib test for now, since it's not reliable */
872 	if (ring == &ring->adev->vce.ring[1])
873 		return 0;
874 
875 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
876 	if (r) {
877 		DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
878 		goto error;
879 	}
880 
881 	r = amdgpu_vce_get_destroy_msg(ring, 1, &fence);
882 	if (r) {
883 		DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
884 		goto error;
885 	}
886 
887 	r = fence_wait(fence, false);
888 	if (r) {
889 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
890 	} else {
891 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
892 	}
893 error:
894 	fence_put(fence);
895 	return r;
896 }
897