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