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 
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	"amdgpu/bonaire_vce.bin"
44 #define FIRMWARE_KABINI	"amdgpu/kabini_vce.bin"
45 #define FIRMWARE_KAVERI	"amdgpu/kaveri_vce.bin"
46 #define FIRMWARE_HAWAII	"amdgpu/hawaii_vce.bin"
47 #define FIRMWARE_MULLINS	"amdgpu/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 #define FIRMWARE_VEGAM		"amdgpu/vegam_vce.bin"
57 
58 #define FIRMWARE_VEGA10		"amdgpu/vega10_vce.bin"
59 #define FIRMWARE_VEGA12		"amdgpu/vega12_vce.bin"
60 #define FIRMWARE_VEGA20		"amdgpu/vega20_vce.bin"
61 
62 #ifdef CONFIG_DRM_AMDGPU_CIK
63 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
64 MODULE_FIRMWARE(FIRMWARE_KABINI);
65 MODULE_FIRMWARE(FIRMWARE_KAVERI);
66 MODULE_FIRMWARE(FIRMWARE_HAWAII);
67 MODULE_FIRMWARE(FIRMWARE_MULLINS);
68 #endif
69 MODULE_FIRMWARE(FIRMWARE_TONGA);
70 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
71 MODULE_FIRMWARE(FIRMWARE_FIJI);
72 MODULE_FIRMWARE(FIRMWARE_STONEY);
73 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
74 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
75 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
76 MODULE_FIRMWARE(FIRMWARE_VEGAM);
77 
78 MODULE_FIRMWARE(FIRMWARE_VEGA10);
79 MODULE_FIRMWARE(FIRMWARE_VEGA12);
80 MODULE_FIRMWARE(FIRMWARE_VEGA20);
81 
82 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
83 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
84 				     struct amdgpu_bo *bo,
85 				     struct dma_fence **fence);
86 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
87 				      bool direct, struct dma_fence **fence);
88 
89 /**
90  * amdgpu_vce_init - allocate memory, load vce firmware
91  *
92  * @adev: amdgpu_device pointer
93  *
94  * First step to get VCE online, allocate memory and load the firmware
95  */
96 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
97 {
98 	const char *fw_name;
99 	const struct common_firmware_header *hdr;
100 	unsigned ucode_version, version_major, version_minor, binary_id;
101 	int i, r;
102 
103 	switch (adev->asic_type) {
104 #ifdef CONFIG_DRM_AMDGPU_CIK
105 	case CHIP_BONAIRE:
106 		fw_name = FIRMWARE_BONAIRE;
107 		break;
108 	case CHIP_KAVERI:
109 		fw_name = FIRMWARE_KAVERI;
110 		break;
111 	case CHIP_KABINI:
112 		fw_name = FIRMWARE_KABINI;
113 		break;
114 	case CHIP_HAWAII:
115 		fw_name = FIRMWARE_HAWAII;
116 		break;
117 	case CHIP_MULLINS:
118 		fw_name = FIRMWARE_MULLINS;
119 		break;
120 #endif
121 	case CHIP_TONGA:
122 		fw_name = FIRMWARE_TONGA;
123 		break;
124 	case CHIP_CARRIZO:
125 		fw_name = FIRMWARE_CARRIZO;
126 		break;
127 	case CHIP_FIJI:
128 		fw_name = FIRMWARE_FIJI;
129 		break;
130 	case CHIP_STONEY:
131 		fw_name = FIRMWARE_STONEY;
132 		break;
133 	case CHIP_POLARIS10:
134 		fw_name = FIRMWARE_POLARIS10;
135 		break;
136 	case CHIP_POLARIS11:
137 		fw_name = FIRMWARE_POLARIS11;
138 		break;
139 	case CHIP_POLARIS12:
140 		fw_name = FIRMWARE_POLARIS12;
141 		break;
142 	case CHIP_VEGAM:
143 		fw_name = FIRMWARE_VEGAM;
144 		break;
145 	case CHIP_VEGA10:
146 		fw_name = FIRMWARE_VEGA10;
147 		break;
148 	case CHIP_VEGA12:
149 		fw_name = FIRMWARE_VEGA12;
150 		break;
151 	case CHIP_VEGA20:
152 		fw_name = FIRMWARE_VEGA20;
153 		break;
154 
155 	default:
156 		return -EINVAL;
157 	}
158 
159 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
160 	if (r) {
161 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
162 			fw_name);
163 		return r;
164 	}
165 
166 	r = amdgpu_ucode_validate(adev->vce.fw);
167 	if (r) {
168 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
169 			fw_name);
170 		release_firmware(adev->vce.fw);
171 		adev->vce.fw = NULL;
172 		return r;
173 	}
174 
175 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
176 
177 	ucode_version = le32_to_cpu(hdr->ucode_version);
178 	version_major = (ucode_version >> 20) & 0xfff;
179 	version_minor = (ucode_version >> 8) & 0xfff;
180 	binary_id = ucode_version & 0xff;
181 	DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
182 		version_major, version_minor, binary_id);
183 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
184 				(binary_id << 8));
185 
186 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
187 				    AMDGPU_GEM_DOMAIN_VRAM, &adev->vce.vcpu_bo,
188 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
189 	if (r) {
190 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
191 		return r;
192 	}
193 
194 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
195 		atomic_set(&adev->vce.handles[i], 0);
196 		adev->vce.filp[i] = NULL;
197 	}
198 
199 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
200 	mutex_init(&adev->vce.idle_mutex);
201 
202 	return 0;
203 }
204 
205 /**
206  * amdgpu_vce_fini - free memory
207  *
208  * @adev: amdgpu_device pointer
209  *
210  * Last step on VCE teardown, free firmware memory
211  */
212 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
213 {
214 	unsigned i;
215 
216 	if (adev->vce.vcpu_bo == NULL)
217 		return 0;
218 
219 	cancel_delayed_work_sync(&adev->vce.idle_work);
220 	drm_sched_entity_destroy(&adev->vce.entity);
221 
222 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
223 		(void **)&adev->vce.cpu_addr);
224 
225 	for (i = 0; i < adev->vce.num_rings; i++)
226 		amdgpu_ring_fini(&adev->vce.ring[i]);
227 
228 	release_firmware(adev->vce.fw);
229 	mutex_destroy(&adev->vce.idle_mutex);
230 
231 	return 0;
232 }
233 
234 /**
235  * amdgpu_vce_entity_init - init entity
236  *
237  * @adev: amdgpu_device pointer
238  *
239  */
240 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
241 {
242 	struct amdgpu_ring *ring;
243 	struct drm_gpu_scheduler *sched;
244 	int r;
245 
246 	ring = &adev->vce.ring[0];
247 	sched = &ring->sched;
248 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
249 				  &sched, 1, NULL);
250 	if (r != 0) {
251 		DRM_ERROR("Failed setting up VCE run queue.\n");
252 		return r;
253 	}
254 
255 	return 0;
256 }
257 
258 /**
259  * amdgpu_vce_suspend - unpin VCE fw memory
260  *
261  * @adev: amdgpu_device pointer
262  *
263  */
264 int amdgpu_vce_suspend(struct amdgpu_device *adev)
265 {
266 	int i;
267 
268 	cancel_delayed_work_sync(&adev->vce.idle_work);
269 
270 	if (adev->vce.vcpu_bo == NULL)
271 		return 0;
272 
273 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
274 		if (atomic_read(&adev->vce.handles[i]))
275 			break;
276 
277 	if (i == AMDGPU_MAX_VCE_HANDLES)
278 		return 0;
279 
280 	/* TODO: suspending running encoding sessions isn't supported */
281 	return -EINVAL;
282 }
283 
284 /**
285  * amdgpu_vce_resume - pin VCE fw memory
286  *
287  * @adev: amdgpu_device pointer
288  *
289  */
290 int amdgpu_vce_resume(struct amdgpu_device *adev)
291 {
292 	void *cpu_addr;
293 	const struct common_firmware_header *hdr;
294 	unsigned offset;
295 	int r;
296 
297 	if (adev->vce.vcpu_bo == NULL)
298 		return -EINVAL;
299 
300 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
301 	if (r) {
302 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
303 		return r;
304 	}
305 
306 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
307 	if (r) {
308 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
309 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
310 		return r;
311 	}
312 
313 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
314 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
315 	memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
316 		    adev->vce.fw->size - offset);
317 
318 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
319 
320 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
321 
322 	return 0;
323 }
324 
325 /**
326  * amdgpu_vce_idle_work_handler - power off VCE
327  *
328  * @work: pointer to work structure
329  *
330  * power of VCE when it's not used any more
331  */
332 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
333 {
334 	struct amdgpu_device *adev =
335 		container_of(work, struct amdgpu_device, vce.idle_work.work);
336 	unsigned i, count = 0;
337 
338 	for (i = 0; i < adev->vce.num_rings; i++)
339 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
340 
341 	if (count == 0) {
342 		if (adev->pm.dpm_enabled) {
343 			amdgpu_dpm_enable_vce(adev, false);
344 		} else {
345 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
346 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
347 							       AMD_PG_STATE_GATE);
348 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
349 							       AMD_CG_STATE_GATE);
350 		}
351 	} else {
352 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
353 	}
354 }
355 
356 /**
357  * amdgpu_vce_ring_begin_use - power up VCE
358  *
359  * @ring: amdgpu ring
360  *
361  * Make sure VCE is powerd up when we want to use it
362  */
363 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
364 {
365 	struct amdgpu_device *adev = ring->adev;
366 	bool set_clocks;
367 
368 	if (amdgpu_sriov_vf(adev))
369 		return;
370 
371 	mutex_lock(&adev->vce.idle_mutex);
372 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
373 	if (set_clocks) {
374 		if (adev->pm.dpm_enabled) {
375 			amdgpu_dpm_enable_vce(adev, true);
376 		} else {
377 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
378 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
379 							       AMD_CG_STATE_UNGATE);
380 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
381 							       AMD_PG_STATE_UNGATE);
382 
383 		}
384 	}
385 	mutex_unlock(&adev->vce.idle_mutex);
386 }
387 
388 /**
389  * amdgpu_vce_ring_end_use - power VCE down
390  *
391  * @ring: amdgpu ring
392  *
393  * Schedule work to power VCE down again
394  */
395 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
396 {
397 	if (!amdgpu_sriov_vf(ring->adev))
398 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
399 }
400 
401 /**
402  * amdgpu_vce_free_handles - free still open VCE handles
403  *
404  * @adev: amdgpu_device pointer
405  * @filp: drm file pointer
406  *
407  * Close all VCE handles still open by this file pointer
408  */
409 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
410 {
411 	struct amdgpu_ring *ring = &adev->vce.ring[0];
412 	int i, r;
413 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
414 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
415 
416 		if (!handle || adev->vce.filp[i] != filp)
417 			continue;
418 
419 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
420 		if (r)
421 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
422 
423 		adev->vce.filp[i] = NULL;
424 		atomic_set(&adev->vce.handles[i], 0);
425 	}
426 }
427 
428 /**
429  * amdgpu_vce_get_create_msg - generate a VCE create msg
430  *
431  * @adev: amdgpu_device pointer
432  * @ring: ring we should submit the msg to
433  * @handle: VCE session handle to use
434  * @fence: optional fence to return
435  *
436  * Open up a stream for HW test
437  */
438 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
439 				     struct amdgpu_bo *bo,
440 				     struct dma_fence **fence)
441 {
442 	const unsigned ib_size_dw = 1024;
443 	struct amdgpu_job *job;
444 	struct amdgpu_ib *ib;
445 	struct dma_fence *f = NULL;
446 	uint64_t addr;
447 	int i, r;
448 
449 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
450 	if (r)
451 		return r;
452 
453 	ib = &job->ibs[0];
454 
455 	addr = amdgpu_bo_gpu_offset(bo);
456 
457 	/* stitch together an VCE create msg */
458 	ib->length_dw = 0;
459 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
460 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
461 	ib->ptr[ib->length_dw++] = handle;
462 
463 	if ((ring->adev->vce.fw_version >> 24) >= 52)
464 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
465 	else
466 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
467 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
468 	ib->ptr[ib->length_dw++] = 0x00000000;
469 	ib->ptr[ib->length_dw++] = 0x00000042;
470 	ib->ptr[ib->length_dw++] = 0x0000000a;
471 	ib->ptr[ib->length_dw++] = 0x00000001;
472 	ib->ptr[ib->length_dw++] = 0x00000080;
473 	ib->ptr[ib->length_dw++] = 0x00000060;
474 	ib->ptr[ib->length_dw++] = 0x00000100;
475 	ib->ptr[ib->length_dw++] = 0x00000100;
476 	ib->ptr[ib->length_dw++] = 0x0000000c;
477 	ib->ptr[ib->length_dw++] = 0x00000000;
478 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
479 		ib->ptr[ib->length_dw++] = 0x00000000;
480 		ib->ptr[ib->length_dw++] = 0x00000000;
481 		ib->ptr[ib->length_dw++] = 0x00000000;
482 		ib->ptr[ib->length_dw++] = 0x00000000;
483 	}
484 
485 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
486 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
487 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
488 	ib->ptr[ib->length_dw++] = addr;
489 	ib->ptr[ib->length_dw++] = 0x00000001;
490 
491 	for (i = ib->length_dw; i < ib_size_dw; ++i)
492 		ib->ptr[i] = 0x0;
493 
494 	r = amdgpu_job_submit_direct(job, ring, &f);
495 	if (r)
496 		goto err;
497 
498 	if (fence)
499 		*fence = dma_fence_get(f);
500 	dma_fence_put(f);
501 	return 0;
502 
503 err:
504 	amdgpu_job_free(job);
505 	return r;
506 }
507 
508 /**
509  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
510  *
511  * @adev: amdgpu_device pointer
512  * @ring: ring we should submit the msg to
513  * @handle: VCE session handle to use
514  * @fence: optional fence to return
515  *
516  * Close up a stream for HW test or if userspace failed to do so
517  */
518 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
519 				      bool direct, struct dma_fence **fence)
520 {
521 	const unsigned ib_size_dw = 1024;
522 	struct amdgpu_job *job;
523 	struct amdgpu_ib *ib;
524 	struct dma_fence *f = NULL;
525 	int i, r;
526 
527 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4, &job);
528 	if (r)
529 		return r;
530 
531 	ib = &job->ibs[0];
532 
533 	/* stitch together an VCE destroy msg */
534 	ib->length_dw = 0;
535 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
536 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
537 	ib->ptr[ib->length_dw++] = handle;
538 
539 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
540 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
541 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
542 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
543 	ib->ptr[ib->length_dw++] = 0x00000000;
544 	ib->ptr[ib->length_dw++] = 0x00000000;
545 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
546 	ib->ptr[ib->length_dw++] = 0x00000000;
547 
548 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
549 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
550 
551 	for (i = ib->length_dw; i < ib_size_dw; ++i)
552 		ib->ptr[i] = 0x0;
553 
554 	if (direct)
555 		r = amdgpu_job_submit_direct(job, ring, &f);
556 	else
557 		r = amdgpu_job_submit(job, &ring->adev->vce.entity,
558 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
559 	if (r)
560 		goto err;
561 
562 	if (fence)
563 		*fence = dma_fence_get(f);
564 	dma_fence_put(f);
565 	return 0;
566 
567 err:
568 	amdgpu_job_free(job);
569 	return r;
570 }
571 
572 /**
573  * amdgpu_vce_cs_validate_bo - make sure not to cross 4GB boundary
574  *
575  * @p: parser context
576  * @lo: address of lower dword
577  * @hi: address of higher dword
578  * @size: minimum size
579  * @index: bs/fb index
580  *
581  * Make sure that no BO cross a 4GB boundary.
582  */
583 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p, uint32_t ib_idx,
584 				  int lo, int hi, unsigned size, int32_t index)
585 {
586 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
587 	struct ttm_operation_ctx ctx = { false, false };
588 	struct amdgpu_bo_va_mapping *mapping;
589 	unsigned i, fpfn, lpfn;
590 	struct amdgpu_bo *bo;
591 	uint64_t addr;
592 	int r;
593 
594 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
595 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
596 	if (index >= 0) {
597 		addr += offset;
598 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
599 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
600 	} else {
601 		fpfn = 0;
602 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
603 	}
604 
605 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
606 	if (r) {
607 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
608 			  addr, lo, hi, size, index);
609 		return r;
610 	}
611 
612 	for (i = 0; i < bo->placement.num_placement; ++i) {
613 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
614 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
615 			min(bo->placements[i].lpfn, lpfn) : lpfn;
616 	}
617 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
618 }
619 
620 
621 /**
622  * amdgpu_vce_cs_reloc - command submission relocation
623  *
624  * @p: parser context
625  * @lo: address of lower dword
626  * @hi: address of higher dword
627  * @size: minimum size
628  *
629  * Patch relocation inside command stream with real buffer address
630  */
631 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
632 			       int lo, int hi, unsigned size, uint32_t index)
633 {
634 	struct amdgpu_bo_va_mapping *mapping;
635 	struct amdgpu_bo *bo;
636 	uint64_t addr;
637 	int r;
638 
639 	if (index == 0xffffffff)
640 		index = 0;
641 
642 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
643 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
644 	addr += ((uint64_t)size) * ((uint64_t)index);
645 
646 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
647 	if (r) {
648 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
649 			  addr, lo, hi, size, index);
650 		return r;
651 	}
652 
653 	if ((addr + (uint64_t)size) >
654 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
655 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
656 			  addr, lo, hi);
657 		return -EINVAL;
658 	}
659 
660 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
661 	addr += amdgpu_bo_gpu_offset(bo);
662 	addr -= ((uint64_t)size) * ((uint64_t)index);
663 
664 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
665 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
666 
667 	return 0;
668 }
669 
670 /**
671  * amdgpu_vce_validate_handle - validate stream handle
672  *
673  * @p: parser context
674  * @handle: handle to validate
675  * @allocated: allocated a new handle?
676  *
677  * Validates the handle and return the found session index or -EINVAL
678  * we we don't have another free session index.
679  */
680 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
681 				      uint32_t handle, uint32_t *allocated)
682 {
683 	unsigned i;
684 
685 	/* validate the handle */
686 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
687 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
688 			if (p->adev->vce.filp[i] != p->filp) {
689 				DRM_ERROR("VCE handle collision detected!\n");
690 				return -EINVAL;
691 			}
692 			return i;
693 		}
694 	}
695 
696 	/* handle not found try to alloc a new one */
697 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
698 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
699 			p->adev->vce.filp[i] = p->filp;
700 			p->adev->vce.img_size[i] = 0;
701 			*allocated |= 1 << i;
702 			return i;
703 		}
704 	}
705 
706 	DRM_ERROR("No more free VCE handles!\n");
707 	return -EINVAL;
708 }
709 
710 /**
711  * amdgpu_vce_cs_parse - parse and validate the command stream
712  *
713  * @p: parser context
714  *
715  */
716 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
717 {
718 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
719 	unsigned fb_idx = 0, bs_idx = 0;
720 	int session_idx = -1;
721 	uint32_t destroyed = 0;
722 	uint32_t created = 0;
723 	uint32_t allocated = 0;
724 	uint32_t tmp, handle = 0;
725 	uint32_t *size = &tmp;
726 	unsigned idx;
727 	int i, r = 0;
728 
729 	p->job->vm = NULL;
730 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
731 
732 	for (idx = 0; idx < ib->length_dw;) {
733 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
734 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
735 
736 		if ((len < 8) || (len & 3)) {
737 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
738 			r = -EINVAL;
739 			goto out;
740 		}
741 
742 		switch (cmd) {
743 		case 0x00000002: /* task info */
744 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
745 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
746 			break;
747 
748 		case 0x03000001: /* encode */
749 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 10,
750 						   idx + 9, 0, 0);
751 			if (r)
752 				goto out;
753 
754 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 12,
755 						   idx + 11, 0, 0);
756 			if (r)
757 				goto out;
758 			break;
759 
760 		case 0x05000001: /* context buffer */
761 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
762 						   idx + 2, 0, 0);
763 			if (r)
764 				goto out;
765 			break;
766 
767 		case 0x05000004: /* video bitstream buffer */
768 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
769 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
770 						   tmp, bs_idx);
771 			if (r)
772 				goto out;
773 			break;
774 
775 		case 0x05000005: /* feedback buffer */
776 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
777 						   4096, fb_idx);
778 			if (r)
779 				goto out;
780 			break;
781 
782 		case 0x0500000d: /* MV buffer */
783 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
784 							idx + 2, 0, 0);
785 			if (r)
786 				goto out;
787 
788 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 8,
789 							idx + 7, 0, 0);
790 			if (r)
791 				goto out;
792 			break;
793 		}
794 
795 		idx += len / 4;
796 	}
797 
798 	for (idx = 0; idx < ib->length_dw;) {
799 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
800 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
801 
802 		switch (cmd) {
803 		case 0x00000001: /* session */
804 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
805 			session_idx = amdgpu_vce_validate_handle(p, handle,
806 								 &allocated);
807 			if (session_idx < 0) {
808 				r = session_idx;
809 				goto out;
810 			}
811 			size = &p->adev->vce.img_size[session_idx];
812 			break;
813 
814 		case 0x00000002: /* task info */
815 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
816 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
817 			break;
818 
819 		case 0x01000001: /* create */
820 			created |= 1 << session_idx;
821 			if (destroyed & (1 << session_idx)) {
822 				destroyed &= ~(1 << session_idx);
823 				allocated |= 1 << session_idx;
824 
825 			} else if (!(allocated & (1 << session_idx))) {
826 				DRM_ERROR("Handle already in use!\n");
827 				r = -EINVAL;
828 				goto out;
829 			}
830 
831 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
832 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
833 				8 * 3 / 2;
834 			break;
835 
836 		case 0x04000001: /* config extension */
837 		case 0x04000002: /* pic control */
838 		case 0x04000005: /* rate control */
839 		case 0x04000007: /* motion estimation */
840 		case 0x04000008: /* rdo */
841 		case 0x04000009: /* vui */
842 		case 0x05000002: /* auxiliary buffer */
843 		case 0x05000009: /* clock table */
844 			break;
845 
846 		case 0x0500000c: /* hw config */
847 			switch (p->adev->asic_type) {
848 #ifdef CONFIG_DRM_AMDGPU_CIK
849 			case CHIP_KAVERI:
850 			case CHIP_MULLINS:
851 #endif
852 			case CHIP_CARRIZO:
853 				break;
854 			default:
855 				r = -EINVAL;
856 				goto out;
857 			}
858 			break;
859 
860 		case 0x03000001: /* encode */
861 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
862 						*size, 0);
863 			if (r)
864 				goto out;
865 
866 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
867 						*size / 3, 0);
868 			if (r)
869 				goto out;
870 			break;
871 
872 		case 0x02000001: /* destroy */
873 			destroyed |= 1 << session_idx;
874 			break;
875 
876 		case 0x05000001: /* context buffer */
877 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
878 						*size * 2, 0);
879 			if (r)
880 				goto out;
881 			break;
882 
883 		case 0x05000004: /* video bitstream buffer */
884 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
885 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
886 						tmp, bs_idx);
887 			if (r)
888 				goto out;
889 			break;
890 
891 		case 0x05000005: /* feedback buffer */
892 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
893 						4096, fb_idx);
894 			if (r)
895 				goto out;
896 			break;
897 
898 		case 0x0500000d: /* MV buffer */
899 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3,
900 							idx + 2, *size, 0);
901 			if (r)
902 				goto out;
903 
904 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 8,
905 							idx + 7, *size / 12, 0);
906 			if (r)
907 				goto out;
908 			break;
909 
910 		default:
911 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
912 			r = -EINVAL;
913 			goto out;
914 		}
915 
916 		if (session_idx == -1) {
917 			DRM_ERROR("no session command at start of IB\n");
918 			r = -EINVAL;
919 			goto out;
920 		}
921 
922 		idx += len / 4;
923 	}
924 
925 	if (allocated & ~created) {
926 		DRM_ERROR("New session without create command!\n");
927 		r = -ENOENT;
928 	}
929 
930 out:
931 	if (!r) {
932 		/* No error, free all destroyed handle slots */
933 		tmp = destroyed;
934 	} else {
935 		/* Error during parsing, free all allocated handle slots */
936 		tmp = allocated;
937 	}
938 
939 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
940 		if (tmp & (1 << i))
941 			atomic_set(&p->adev->vce.handles[i], 0);
942 
943 	return r;
944 }
945 
946 /**
947  * amdgpu_vce_cs_parse_vm - parse the command stream in VM mode
948  *
949  * @p: parser context
950  *
951  */
952 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p, uint32_t ib_idx)
953 {
954 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
955 	int session_idx = -1;
956 	uint32_t destroyed = 0;
957 	uint32_t created = 0;
958 	uint32_t allocated = 0;
959 	uint32_t tmp, handle = 0;
960 	int i, r = 0, idx = 0;
961 
962 	while (idx < ib->length_dw) {
963 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
964 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
965 
966 		if ((len < 8) || (len & 3)) {
967 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
968 			r = -EINVAL;
969 			goto out;
970 		}
971 
972 		switch (cmd) {
973 		case 0x00000001: /* session */
974 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
975 			session_idx = amdgpu_vce_validate_handle(p, handle,
976 								 &allocated);
977 			if (session_idx < 0) {
978 				r = session_idx;
979 				goto out;
980 			}
981 			break;
982 
983 		case 0x01000001: /* create */
984 			created |= 1 << session_idx;
985 			if (destroyed & (1 << session_idx)) {
986 				destroyed &= ~(1 << session_idx);
987 				allocated |= 1 << session_idx;
988 
989 			} else if (!(allocated & (1 << session_idx))) {
990 				DRM_ERROR("Handle already in use!\n");
991 				r = -EINVAL;
992 				goto out;
993 			}
994 
995 			break;
996 
997 		case 0x02000001: /* destroy */
998 			destroyed |= 1 << session_idx;
999 			break;
1000 
1001 		default:
1002 			break;
1003 		}
1004 
1005 		if (session_idx == -1) {
1006 			DRM_ERROR("no session command at start of IB\n");
1007 			r = -EINVAL;
1008 			goto out;
1009 		}
1010 
1011 		idx += len / 4;
1012 	}
1013 
1014 	if (allocated & ~created) {
1015 		DRM_ERROR("New session without create command!\n");
1016 		r = -ENOENT;
1017 	}
1018 
1019 out:
1020 	if (!r) {
1021 		/* No error, free all destroyed handle slots */
1022 		tmp = destroyed;
1023 		amdgpu_ib_free(p->adev, ib, NULL);
1024 	} else {
1025 		/* Error during parsing, free all allocated handle slots */
1026 		tmp = allocated;
1027 	}
1028 
1029 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1030 		if (tmp & (1 << i))
1031 			atomic_set(&p->adev->vce.handles[i], 0);
1032 
1033 	return r;
1034 }
1035 
1036 /**
1037  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1038  *
1039  * @ring: engine to use
1040  * @ib: the IB to execute
1041  *
1042  */
1043 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1044 				struct amdgpu_job *job,
1045 				struct amdgpu_ib *ib,
1046 				uint32_t flags)
1047 {
1048 	amdgpu_ring_write(ring, VCE_CMD_IB);
1049 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1050 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1051 	amdgpu_ring_write(ring, ib->length_dw);
1052 }
1053 
1054 /**
1055  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1056  *
1057  * @ring: engine to use
1058  * @fence: the fence
1059  *
1060  */
1061 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1062 				unsigned flags)
1063 {
1064 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1065 
1066 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1067 	amdgpu_ring_write(ring, addr);
1068 	amdgpu_ring_write(ring, upper_32_bits(addr));
1069 	amdgpu_ring_write(ring, seq);
1070 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1071 	amdgpu_ring_write(ring, VCE_CMD_END);
1072 }
1073 
1074 /**
1075  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1076  *
1077  * @ring: the engine to test on
1078  *
1079  */
1080 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1081 {
1082 	struct amdgpu_device *adev = ring->adev;
1083 	uint32_t rptr;
1084 	unsigned i;
1085 	int r, timeout = adev->usec_timeout;
1086 
1087 	/* skip ring test for sriov*/
1088 	if (amdgpu_sriov_vf(adev))
1089 		return 0;
1090 
1091 	r = amdgpu_ring_alloc(ring, 16);
1092 	if (r)
1093 		return r;
1094 
1095 	rptr = amdgpu_ring_get_rptr(ring);
1096 
1097 	amdgpu_ring_write(ring, VCE_CMD_END);
1098 	amdgpu_ring_commit(ring);
1099 
1100 	for (i = 0; i < timeout; i++) {
1101 		if (amdgpu_ring_get_rptr(ring) != rptr)
1102 			break;
1103 		udelay(1);
1104 	}
1105 
1106 	if (i >= timeout)
1107 		r = -ETIMEDOUT;
1108 
1109 	return r;
1110 }
1111 
1112 /**
1113  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1114  *
1115  * @ring: the engine to test on
1116  *
1117  */
1118 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1119 {
1120 	struct dma_fence *fence = NULL;
1121 	struct amdgpu_bo *bo = NULL;
1122 	long r;
1123 
1124 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1125 	if (ring != &ring->adev->vce.ring[0])
1126 		return 0;
1127 
1128 	r = amdgpu_bo_create_reserved(ring->adev, 512, PAGE_SIZE,
1129 				      AMDGPU_GEM_DOMAIN_VRAM,
1130 				      &bo, NULL, NULL);
1131 	if (r)
1132 		return r;
1133 
1134 	r = amdgpu_vce_get_create_msg(ring, 1, bo, NULL);
1135 	if (r)
1136 		goto error;
1137 
1138 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1139 	if (r)
1140 		goto error;
1141 
1142 	r = dma_fence_wait_timeout(fence, false, timeout);
1143 	if (r == 0)
1144 		r = -ETIMEDOUT;
1145 	else if (r > 0)
1146 		r = 0;
1147 
1148 error:
1149 	dma_fence_put(fence);
1150 	amdgpu_bo_unreserve(bo);
1151 	amdgpu_bo_unref(&bo);
1152 	return r;
1153 }
1154