xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_vce.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: amdgpu_vce.c,v 1.5 2020/02/14 14:34:58 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 1.5 2020/02/14 14:34:58 maya Exp $");
32 
33 #include <linux/firmware.h>
34 #include <linux/module.h>
35 #include <drm/drmP.h>
36 #include <drm/drm.h>
37 
38 #include "amdgpu.h"
39 #include "amdgpu_pm.h"
40 #include "amdgpu_vce.h"
41 #include "cikd.h"
42 
43 /* 1 second timeout */
44 #define VCE_IDLE_TIMEOUT_MS	1000
45 
46 /* Firmware Names */
47 #ifdef CONFIG_DRM_AMDGPU_CIK
48 #define FIRMWARE_BONAIRE	"radeon/bonaire_vce.bin"
49 #define FIRMWARE_KABINI 	"radeon/kabini_vce.bin"
50 #define FIRMWARE_KAVERI 	"radeon/kaveri_vce.bin"
51 #define FIRMWARE_HAWAII 	"radeon/hawaii_vce.bin"
52 #define FIRMWARE_MULLINS	"radeon/mullins_vce.bin"
53 #endif
54 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
55 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
56 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
57 #define FIRMWARE_STONEY		"amdgpu/stoney_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 
71 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
72 
73 /**
74  * amdgpu_vce_init - allocate memory, load vce firmware
75  *
76  * @adev: amdgpu_device pointer
77  *
78  * First step to get VCE online, allocate memory and load the firmware
79  */
80 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
81 {
82 	const char *fw_name;
83 	const struct common_firmware_header *hdr;
84 	unsigned ucode_version, version_major, version_minor, binary_id;
85 	int i, r;
86 
87 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
88 
89 	switch (adev->asic_type) {
90 #ifdef CONFIG_DRM_AMDGPU_CIK
91 	case CHIP_BONAIRE:
92 		fw_name = FIRMWARE_BONAIRE;
93 		break;
94 	case CHIP_KAVERI:
95 		fw_name = FIRMWARE_KAVERI;
96 		break;
97 	case CHIP_KABINI:
98 		fw_name = FIRMWARE_KABINI;
99 		break;
100 	case CHIP_HAWAII:
101 		fw_name = FIRMWARE_HAWAII;
102 		break;
103 	case CHIP_MULLINS:
104 		fw_name = FIRMWARE_MULLINS;
105 		break;
106 #endif
107 	case CHIP_TONGA:
108 		fw_name = FIRMWARE_TONGA;
109 		break;
110 	case CHIP_CARRIZO:
111 		fw_name = FIRMWARE_CARRIZO;
112 		break;
113 	case CHIP_FIJI:
114 		fw_name = FIRMWARE_FIJI;
115 		break;
116 	case CHIP_STONEY:
117 		fw_name = FIRMWARE_STONEY;
118 		break;
119 
120 	default:
121 		return -EINVAL;
122 	}
123 
124 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
125 	if (r) {
126 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
127 			fw_name);
128 		return r;
129 	}
130 
131 	r = amdgpu_ucode_validate(adev->vce.fw);
132 	if (r) {
133 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
134 			fw_name);
135 		release_firmware(adev->vce.fw);
136 		adev->vce.fw = NULL;
137 		return r;
138 	}
139 
140 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
141 
142 	ucode_version = le32_to_cpu(hdr->ucode_version);
143 	version_major = (ucode_version >> 20) & 0xfff;
144 	version_minor = (ucode_version >> 8) & 0xfff;
145 	binary_id = ucode_version & 0xff;
146 	DRM_INFO("Found VCE firmware Version: %x.%x Binary ID: %x\n",
147 		version_major, version_minor, binary_id);
148 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
149 				(binary_id << 8));
150 
151 	/* allocate firmware, stack and heap BO */
152 
153 	r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
154 			     AMDGPU_GEM_DOMAIN_VRAM,
155 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
156 			     NULL, NULL, &adev->vce.vcpu_bo);
157 	if (r) {
158 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
159 		return r;
160 	}
161 
162 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
163 	if (r) {
164 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
165 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
166 		return r;
167 	}
168 
169 	r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
170 			  &adev->vce.gpu_addr);
171 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
172 	if (r) {
173 		amdgpu_bo_unref(&adev->vce.vcpu_bo);
174 		dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
175 		return r;
176 	}
177 
178 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
179 		atomic_set(&adev->vce.handles[i], 0);
180 		adev->vce.filp[i] = NULL;
181 	}
182 
183 	return 0;
184 }
185 
186 /**
187  * amdgpu_vce_fini - free memory
188  *
189  * @adev: amdgpu_device pointer
190  *
191  * Last step on VCE teardown, free firmware memory
192  */
193 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
194 {
195 	if (adev->vce.vcpu_bo == NULL)
196 		return 0;
197 
198 	amdgpu_bo_unref(&adev->vce.vcpu_bo);
199 
200 	amdgpu_ring_fini(&adev->vce.ring[0]);
201 	amdgpu_ring_fini(&adev->vce.ring[1]);
202 
203 	release_firmware(adev->vce.fw);
204 
205 	return 0;
206 }
207 
208 /**
209  * amdgpu_vce_suspend - unpin VCE fw memory
210  *
211  * @adev: amdgpu_device pointer
212  *
213  */
214 int amdgpu_vce_suspend(struct amdgpu_device *adev)
215 {
216 	int i;
217 
218 	if (adev->vce.vcpu_bo == NULL)
219 		return 0;
220 
221 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
222 		if (atomic_read(&adev->vce.handles[i]))
223 			break;
224 
225 	if (i == AMDGPU_MAX_VCE_HANDLES)
226 		return 0;
227 
228 	cancel_delayed_work_sync(&adev->vce.idle_work);
229 	/* TODO: suspending running encoding sessions isn't supported */
230 	return -EINVAL;
231 }
232 
233 /**
234  * amdgpu_vce_resume - pin VCE fw memory
235  *
236  * @adev: amdgpu_device pointer
237  *
238  */
239 int amdgpu_vce_resume(struct amdgpu_device *adev)
240 {
241 	void *cpu_addr;
242 	const struct common_firmware_header *hdr;
243 	unsigned offset;
244 	int r;
245 
246 	if (adev->vce.vcpu_bo == NULL)
247 		return -EINVAL;
248 
249 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
250 	if (r) {
251 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
252 		return r;
253 	}
254 
255 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
256 	if (r) {
257 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
258 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
259 		return r;
260 	}
261 
262 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
263 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
264 	memcpy(cpu_addr, (adev->vce.fw->data) + offset,
265 		(adev->vce.fw->size) - offset);
266 
267 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
268 
269 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
270 
271 	return 0;
272 }
273 
274 /**
275  * amdgpu_vce_idle_work_handler - power off VCE
276  *
277  * @work: pointer to work structure
278  *
279  * power of VCE when it's not used any more
280  */
281 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
282 {
283 	struct amdgpu_device *adev =
284 		container_of(work, struct amdgpu_device, vce.idle_work.work);
285 
286 	if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
287 	    (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
288 		if (adev->pm.dpm_enabled) {
289 			amdgpu_dpm_enable_vce(adev, false);
290 		} else {
291 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
292 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
293 							    AMD_PG_STATE_GATE);
294 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
295 							    AMD_CG_STATE_GATE);
296 		}
297 	} else {
298 		schedule_delayed_work(&adev->vce.idle_work,
299 				      msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
300 	}
301 }
302 
303 /**
304  * amdgpu_vce_note_usage - power up VCE
305  *
306  * @adev: amdgpu_device pointer
307  *
308  * Make sure VCE is powerd up when we want to use it
309  */
310 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
311 {
312 	bool streams_changed = false;
313 	bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
314 	set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
315 					    msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
316 
317 	if (adev->pm.dpm_enabled) {
318 		/* XXX figure out if the streams changed */
319 		streams_changed = false;
320 	}
321 
322 	if (set_clocks || streams_changed) {
323 		if (adev->pm.dpm_enabled) {
324 			amdgpu_dpm_enable_vce(adev, true);
325 		} else {
326 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
327 			amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
328 							    AMD_CG_STATE_UNGATE);
329 			amdgpu_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
330 							    AMD_PG_STATE_UNGATE);
331 
332 		}
333 	}
334 }
335 
336 /**
337  * amdgpu_vce_free_handles - free still open VCE handles
338  *
339  * @adev: amdgpu_device pointer
340  * @filp: drm file pointer
341  *
342  * Close all VCE handles still open by this file pointer
343  */
344 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
345 {
346 	struct amdgpu_ring *ring = &adev->vce.ring[0];
347 	int i, r;
348 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
349 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
350 		if (!handle || adev->vce.filp[i] != filp)
351 			continue;
352 
353 		amdgpu_vce_note_usage(adev);
354 
355 		r = amdgpu_vce_get_destroy_msg(ring, handle, NULL);
356 		if (r)
357 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
358 
359 		adev->vce.filp[i] = NULL;
360 		atomic_set(&adev->vce.handles[i], 0);
361 	}
362 }
363 
364 static int amdgpu_vce_free_job(
365 	struct amdgpu_job *job)
366 {
367 	amdgpu_ib_free(job->adev, job->ibs);
368 	kfree(job->ibs);
369 	return 0;
370 }
371 
372 /**
373  * amdgpu_vce_get_create_msg - generate a VCE create msg
374  *
375  * @adev: amdgpu_device pointer
376  * @ring: ring we should submit the msg to
377  * @handle: VCE session handle to use
378  * @fence: optional fence to return
379  *
380  * Open up a stream for HW test
381  */
382 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
383 			      struct fence **fence)
384 {
385 	const unsigned ib_size_dw = 1024;
386 	struct amdgpu_ib *ib = NULL;
387 	struct fence *f = NULL;
388 	struct amdgpu_device *adev = ring->adev;
389 	uint64_t dummy;
390 	int i, r;
391 
392 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
393 	if (!ib)
394 		return -ENOMEM;
395 	r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
396 	if (r) {
397 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
398 		kfree(ib);
399 		return r;
400 	}
401 
402 	dummy = ib->gpu_addr + 1024;
403 
404 	/* stitch together an VCE create msg */
405 	ib->length_dw = 0;
406 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
407 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
408 	ib->ptr[ib->length_dw++] = handle;
409 
410 	if ((ring->adev->vce.fw_version >> 24) >= 52)
411 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
412 	else
413 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
414 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
415 	ib->ptr[ib->length_dw++] = 0x00000000;
416 	ib->ptr[ib->length_dw++] = 0x00000042;
417 	ib->ptr[ib->length_dw++] = 0x0000000a;
418 	ib->ptr[ib->length_dw++] = 0x00000001;
419 	ib->ptr[ib->length_dw++] = 0x00000080;
420 	ib->ptr[ib->length_dw++] = 0x00000060;
421 	ib->ptr[ib->length_dw++] = 0x00000100;
422 	ib->ptr[ib->length_dw++] = 0x00000100;
423 	ib->ptr[ib->length_dw++] = 0x0000000c;
424 	ib->ptr[ib->length_dw++] = 0x00000000;
425 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
426 		ib->ptr[ib->length_dw++] = 0x00000000;
427 		ib->ptr[ib->length_dw++] = 0x00000000;
428 		ib->ptr[ib->length_dw++] = 0x00000000;
429 		ib->ptr[ib->length_dw++] = 0x00000000;
430 	}
431 
432 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
433 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
434 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
435 	ib->ptr[ib->length_dw++] = dummy;
436 	ib->ptr[ib->length_dw++] = 0x00000001;
437 
438 	for (i = ib->length_dw; i < ib_size_dw; ++i)
439 		ib->ptr[i] = 0x0;
440 
441 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
442 						 &amdgpu_vce_free_job,
443 						 AMDGPU_FENCE_OWNER_UNDEFINED,
444 						 &f);
445 	if (r)
446 		goto err;
447 	if (fence)
448 		*fence = fence_get(f);
449 	fence_put(f);
450 	if (amdgpu_enable_scheduler)
451 		return 0;
452 err:
453 	amdgpu_ib_free(adev, ib);
454 	kfree(ib);
455 	return r;
456 }
457 
458 /**
459  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
460  *
461  * @adev: amdgpu_device pointer
462  * @ring: ring we should submit the msg to
463  * @handle: VCE session handle to use
464  * @fence: optional fence to return
465  *
466  * Close up a stream for HW test or if userspace failed to do so
467  */
468 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
469 			       struct fence **fence)
470 {
471 	const unsigned ib_size_dw = 1024;
472 	struct amdgpu_ib *ib = NULL;
473 	struct fence *f = NULL;
474 	struct amdgpu_device *adev = ring->adev;
475 	uint64_t dummy;
476 	int i, r;
477 
478 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
479 	if (!ib)
480 		return -ENOMEM;
481 
482 	r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, ib);
483 	if (r) {
484 		kfree(ib);
485 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
486 		return r;
487 	}
488 
489 	dummy = ib->gpu_addr + 1024;
490 
491 	/* stitch together an VCE destroy msg */
492 	ib->length_dw = 0;
493 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
494 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
495 	ib->ptr[ib->length_dw++] = handle;
496 
497 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
498 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
499 	ib->ptr[ib->length_dw++] = upper_32_bits(dummy);
500 	ib->ptr[ib->length_dw++] = dummy;
501 	ib->ptr[ib->length_dw++] = 0x00000001;
502 
503 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
504 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
505 
506 	for (i = ib->length_dw; i < ib_size_dw; ++i)
507 		ib->ptr[i] = 0x0;
508 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
509 						 &amdgpu_vce_free_job,
510 						 AMDGPU_FENCE_OWNER_UNDEFINED,
511 						 &f);
512 	if (r)
513 		goto err;
514 	if (fence)
515 		*fence = fence_get(f);
516 	fence_put(f);
517 	if (amdgpu_enable_scheduler)
518 		return 0;
519 err:
520 	amdgpu_ib_free(adev, ib);
521 	kfree(ib);
522 	return r;
523 }
524 
525 /**
526  * amdgpu_vce_cs_reloc - command submission relocation
527  *
528  * @p: parser context
529  * @lo: address of lower dword
530  * @hi: address of higher dword
531  * @size: minimum size
532  *
533  * Patch relocation inside command stream with real buffer address
534  */
535 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
536 			       int lo, int hi, unsigned size, uint32_t index)
537 {
538 	struct amdgpu_bo_va_mapping *mapping;
539 	struct amdgpu_ib *ib = &p->ibs[ib_idx];
540 	struct amdgpu_bo *bo;
541 	uint64_t addr;
542 
543 	if (index == 0xffffffff)
544 		index = 0;
545 
546 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
547 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
548 	addr += ((uint64_t)size) * ((uint64_t)index);
549 
550 	mapping = amdgpu_cs_find_mapping(p, addr, &bo);
551 	if (mapping == NULL) {
552 		DRM_ERROR("Can't find BO for addr 0x%010"PRIx64" %d %d %d %d\n",
553 			  addr, lo, hi, size, index);
554 		return -EINVAL;
555 	}
556 
557 	if ((addr + (uint64_t)size) >
558 	    ((uint64_t)mapping->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
559 		DRM_ERROR("BO to small for addr 0x%010"PRIx64" %d %d\n",
560 			  addr, lo, hi);
561 		return -EINVAL;
562 	}
563 
564 	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
565 	addr += amdgpu_bo_gpu_offset(bo);
566 	addr -= ((uint64_t)size) * ((uint64_t)index);
567 
568 	ib->ptr[lo] = addr & 0xFFFFFFFF;
569 	ib->ptr[hi] = addr >> 32;
570 
571 	return 0;
572 }
573 
574 /**
575  * amdgpu_vce_validate_handle - validate stream handle
576  *
577  * @p: parser context
578  * @handle: handle to validate
579  * @allocated: allocated a new handle?
580  *
581  * Validates the handle and return the found session index or -EINVAL
582  * we we don't have another free session index.
583  */
584 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
585 				      uint32_t handle, bool *allocated)
586 {
587 	unsigned i;
588 
589 	*allocated = false;
590 
591 	/* validate the handle */
592 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
593 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
594 			if (p->adev->vce.filp[i] != p->filp) {
595 				DRM_ERROR("VCE handle collision detected!\n");
596 				return -EINVAL;
597 			}
598 			return i;
599 		}
600 	}
601 
602 	/* handle not found try to alloc a new one */
603 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
604 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
605 			p->adev->vce.filp[i] = p->filp;
606 			p->adev->vce.img_size[i] = 0;
607 			*allocated = true;
608 			return i;
609 		}
610 	}
611 
612 	DRM_ERROR("No more free VCE handles!\n");
613 	return -EINVAL;
614 }
615 
616 /**
617  * amdgpu_vce_cs_parse - parse and validate the command stream
618  *
619  * @p: parser context
620  *
621  */
622 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
623 {
624 	struct amdgpu_ib *ib = &p->ibs[ib_idx];
625 	unsigned fb_idx = 0, bs_idx = 0;
626 	int session_idx = -1;
627 	bool destroyed = false;
628 	bool created = false;
629 	bool allocated = false;
630 	uint32_t tmp, handle = 0;
631 	uint32_t *size = &tmp;
632 	int i, r = 0, idx = 0;
633 
634 	amdgpu_vce_note_usage(p->adev);
635 
636 	while (idx < ib->length_dw) {
637 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
638 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
639 
640 		if ((len < 8) || (len & 3)) {
641 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
642 			r = -EINVAL;
643 			goto out;
644 		}
645 
646 		if (destroyed) {
647 			DRM_ERROR("No other command allowed after destroy!\n");
648 			r = -EINVAL;
649 			goto out;
650 		}
651 
652 		switch (cmd) {
653 		case 0x00000001: // session
654 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
655 			session_idx = amdgpu_vce_validate_handle(p, handle,
656 								 &allocated);
657 			if (session_idx < 0)
658 				return session_idx;
659 			size = &p->adev->vce.img_size[session_idx];
660 			break;
661 
662 		case 0x00000002: // task info
663 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
664 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
665 			break;
666 
667 		case 0x01000001: // create
668 			created = true;
669 			if (!allocated) {
670 				DRM_ERROR("Handle already in use!\n");
671 				r = -EINVAL;
672 				goto out;
673 			}
674 
675 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
676 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
677 				8 * 3 / 2;
678 			break;
679 
680 		case 0x04000001: // config extension
681 		case 0x04000002: // pic control
682 		case 0x04000005: // rate control
683 		case 0x04000007: // motion estimation
684 		case 0x04000008: // rdo
685 		case 0x04000009: // vui
686 		case 0x05000002: // auxiliary buffer
687 			break;
688 
689 		case 0x03000001: // encode
690 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
691 						*size, 0);
692 			if (r)
693 				goto out;
694 
695 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
696 						*size / 3, 0);
697 			if (r)
698 				goto out;
699 			break;
700 
701 		case 0x02000001: // destroy
702 			destroyed = true;
703 			break;
704 
705 		case 0x05000001: // context buffer
706 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
707 						*size * 2, 0);
708 			if (r)
709 				goto out;
710 			break;
711 
712 		case 0x05000004: // video bitstream buffer
713 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
714 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
715 						tmp, bs_idx);
716 			if (r)
717 				goto out;
718 			break;
719 
720 		case 0x05000005: // feedback buffer
721 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
722 						4096, fb_idx);
723 			if (r)
724 				goto out;
725 			break;
726 
727 		default:
728 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
729 			r = -EINVAL;
730 			goto out;
731 		}
732 
733 		if (session_idx == -1) {
734 			DRM_ERROR("no session command at start of IB\n");
735 			r = -EINVAL;
736 			goto out;
737 		}
738 
739 		idx += len / 4;
740 	}
741 
742 	if (allocated && !created) {
743 		DRM_ERROR("New session without create command!\n");
744 		r = -ENOENT;
745 	}
746 
747 out:
748 	if ((!r && destroyed) || (r && allocated)) {
749 		/*
750 		 * IB contains a destroy msg or we have allocated an
751 		 * handle and got an error, anyway free the handle
752 		 */
753 		for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
754 			atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
755 	}
756 
757 	return r;
758 }
759 
760 /**
761  * amdgpu_vce_ring_emit_semaphore - emit a semaphore command
762  *
763  * @ring: engine to use
764  * @semaphore: address of semaphore
765  * @emit_wait: true=emit wait, false=emit signal
766  *
767  */
768 bool amdgpu_vce_ring_emit_semaphore(struct amdgpu_ring *ring,
769 				    struct amdgpu_semaphore *semaphore,
770 				    bool emit_wait)
771 {
772 	uint64_t addr = semaphore->gpu_addr;
773 
774 	amdgpu_ring_write(ring, VCE_CMD_SEMAPHORE);
775 	amdgpu_ring_write(ring, (addr >> 3) & 0x000FFFFF);
776 	amdgpu_ring_write(ring, (addr >> 23) & 0x000FFFFF);
777 	amdgpu_ring_write(ring, 0x01003000 | (emit_wait ? 1 : 0));
778 	if (!emit_wait)
779 		amdgpu_ring_write(ring, VCE_CMD_END);
780 
781 	return true;
782 }
783 
784 /**
785  * amdgpu_vce_ring_emit_ib - execute indirect buffer
786  *
787  * @ring: engine to use
788  * @ib: the IB to execute
789  *
790  */
791 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
792 {
793 	amdgpu_ring_write(ring, VCE_CMD_IB);
794 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
795 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
796 	amdgpu_ring_write(ring, ib->length_dw);
797 }
798 
799 /**
800  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
801  *
802  * @ring: engine to use
803  * @fence: the fence
804  *
805  */
806 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
807 				unsigned flags)
808 {
809 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
810 
811 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
812 	amdgpu_ring_write(ring, addr);
813 	amdgpu_ring_write(ring, upper_32_bits(addr));
814 	amdgpu_ring_write(ring, seq);
815 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
816 	amdgpu_ring_write(ring, VCE_CMD_END);
817 }
818 
819 /**
820  * amdgpu_vce_ring_test_ring - test if VCE ring is working
821  *
822  * @ring: the engine to test on
823  *
824  */
825 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
826 {
827 	struct amdgpu_device *adev = ring->adev;
828 	uint32_t rptr = amdgpu_ring_get_rptr(ring);
829 	unsigned i;
830 	int r;
831 
832 	r = amdgpu_ring_lock(ring, 16);
833 	if (r) {
834 		DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
835 			  ring->idx, r);
836 		return r;
837 	}
838 	amdgpu_ring_write(ring, VCE_CMD_END);
839 	amdgpu_ring_unlock_commit(ring);
840 
841 	for (i = 0; i < adev->usec_timeout; i++) {
842 		if (amdgpu_ring_get_rptr(ring) != rptr)
843 			break;
844 		DRM_UDELAY(1);
845 	}
846 
847 	if (i < adev->usec_timeout) {
848 		DRM_INFO("ring test on %d succeeded in %d usecs\n",
849 			 ring->idx, i);
850 	} else {
851 		DRM_ERROR("amdgpu: ring %d test failed\n",
852 			  ring->idx);
853 		r = -ETIMEDOUT;
854 	}
855 
856 	return r;
857 }
858 
859 /**
860  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
861  *
862  * @ring: the engine to test on
863  *
864  */
865 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
866 {
867 	struct fence *fence = NULL;
868 	int r;
869 
870 	/* skip vce ring1 ib test for now, since it's not reliable */
871 	if (ring == &ring->adev->vce.ring[1])
872 		return 0;
873 
874 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
875 	if (r) {
876 		DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
877 		goto error;
878 	}
879 
880 	r = amdgpu_vce_get_destroy_msg(ring, 1, &fence);
881 	if (r) {
882 		DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
883 		goto error;
884 	}
885 
886 	r = fence_wait(fence, false);
887 	if (r) {
888 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
889 	} else {
890 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
891 	}
892 error:
893 	fence_put(fence);
894 	return r;
895 }
896