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  * @size: size for the new BO
94  *
95  * First step to get VCE online, allocate memory and load the firmware
96  */
97 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
98 {
99 	const char *fw_name;
100 	const struct common_firmware_header *hdr;
101 	unsigned ucode_version, version_major, version_minor, binary_id;
102 	int i, r;
103 
104 	switch (adev->asic_type) {
105 #ifdef CONFIG_DRM_AMDGPU_CIK
106 	case CHIP_BONAIRE:
107 		fw_name = FIRMWARE_BONAIRE;
108 		break;
109 	case CHIP_KAVERI:
110 		fw_name = FIRMWARE_KAVERI;
111 		break;
112 	case CHIP_KABINI:
113 		fw_name = FIRMWARE_KABINI;
114 		break;
115 	case CHIP_HAWAII:
116 		fw_name = FIRMWARE_HAWAII;
117 		break;
118 	case CHIP_MULLINS:
119 		fw_name = FIRMWARE_MULLINS;
120 		break;
121 #endif
122 	case CHIP_TONGA:
123 		fw_name = FIRMWARE_TONGA;
124 		break;
125 	case CHIP_CARRIZO:
126 		fw_name = FIRMWARE_CARRIZO;
127 		break;
128 	case CHIP_FIJI:
129 		fw_name = FIRMWARE_FIJI;
130 		break;
131 	case CHIP_STONEY:
132 		fw_name = FIRMWARE_STONEY;
133 		break;
134 	case CHIP_POLARIS10:
135 		fw_name = FIRMWARE_POLARIS10;
136 		break;
137 	case CHIP_POLARIS11:
138 		fw_name = FIRMWARE_POLARIS11;
139 		break;
140 	case CHIP_POLARIS12:
141 		fw_name = FIRMWARE_POLARIS12;
142 		break;
143 	case CHIP_VEGAM:
144 		fw_name = FIRMWARE_VEGAM;
145 		break;
146 	case CHIP_VEGA10:
147 		fw_name = FIRMWARE_VEGA10;
148 		break;
149 	case CHIP_VEGA12:
150 		fw_name = FIRMWARE_VEGA12;
151 		break;
152 	case CHIP_VEGA20:
153 		fw_name = FIRMWARE_VEGA20;
154 		break;
155 
156 	default:
157 		return -EINVAL;
158 	}
159 
160 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
161 	if (r) {
162 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
163 			fw_name);
164 		return r;
165 	}
166 
167 	r = amdgpu_ucode_validate(adev->vce.fw);
168 	if (r) {
169 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
170 			fw_name);
171 		release_firmware(adev->vce.fw);
172 		adev->vce.fw = NULL;
173 		return r;
174 	}
175 
176 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
177 
178 	ucode_version = le32_to_cpu(hdr->ucode_version);
179 	version_major = (ucode_version >> 20) & 0xfff;
180 	version_minor = (ucode_version >> 8) & 0xfff;
181 	binary_id = ucode_version & 0xff;
182 	DRM_INFO("Found VCE firmware Version: %d.%d Binary ID: %d\n",
183 		version_major, version_minor, binary_id);
184 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
185 				(binary_id << 8));
186 
187 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
188 				    AMDGPU_GEM_DOMAIN_VRAM, &adev->vce.vcpu_bo,
189 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
190 	if (r) {
191 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
192 		return r;
193 	}
194 
195 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
196 		atomic_set(&adev->vce.handles[i], 0);
197 		adev->vce.filp[i] = NULL;
198 	}
199 
200 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
201 	mutex_init(&adev->vce.idle_mutex);
202 
203 	return 0;
204 }
205 
206 /**
207  * amdgpu_vce_fini - free memory
208  *
209  * @adev: amdgpu_device pointer
210  *
211  * Last step on VCE teardown, free firmware memory
212  */
213 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
214 {
215 	unsigned i;
216 
217 	if (adev->vce.vcpu_bo == NULL)
218 		return 0;
219 
220 	cancel_delayed_work_sync(&adev->vce.idle_work);
221 	drm_sched_entity_destroy(&adev->vce.entity);
222 
223 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
224 		(void **)&adev->vce.cpu_addr);
225 
226 	for (i = 0; i < adev->vce.num_rings; i++)
227 		amdgpu_ring_fini(&adev->vce.ring[i]);
228 
229 	release_firmware(adev->vce.fw);
230 	mutex_destroy(&adev->vce.idle_mutex);
231 
232 	return 0;
233 }
234 
235 /**
236  * amdgpu_vce_entity_init - init entity
237  *
238  * @adev: amdgpu_device pointer
239  *
240  */
241 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
242 {
243 	struct amdgpu_ring *ring;
244 	struct drm_gpu_scheduler *sched;
245 	int r;
246 
247 	ring = &adev->vce.ring[0];
248 	sched = &ring->sched;
249 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
250 				  &sched, 1, NULL);
251 	if (r != 0) {
252 		DRM_ERROR("Failed setting up VCE run queue.\n");
253 		return r;
254 	}
255 
256 	return 0;
257 }
258 
259 /**
260  * amdgpu_vce_suspend - unpin VCE fw memory
261  *
262  * @adev: amdgpu_device pointer
263  *
264  */
265 int amdgpu_vce_suspend(struct amdgpu_device *adev)
266 {
267 	int i;
268 
269 	cancel_delayed_work_sync(&adev->vce.idle_work);
270 
271 	if (adev->vce.vcpu_bo == NULL)
272 		return 0;
273 
274 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
275 		if (atomic_read(&adev->vce.handles[i]))
276 			break;
277 
278 	if (i == AMDGPU_MAX_VCE_HANDLES)
279 		return 0;
280 
281 	/* TODO: suspending running encoding sessions isn't supported */
282 	return -EINVAL;
283 }
284 
285 /**
286  * amdgpu_vce_resume - pin VCE fw memory
287  *
288  * @adev: amdgpu_device pointer
289  *
290  */
291 int amdgpu_vce_resume(struct amdgpu_device *adev)
292 {
293 	void *cpu_addr;
294 	const struct common_firmware_header *hdr;
295 	unsigned offset;
296 	int r;
297 
298 	if (adev->vce.vcpu_bo == NULL)
299 		return -EINVAL;
300 
301 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
302 	if (r) {
303 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
304 		return r;
305 	}
306 
307 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
308 	if (r) {
309 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
310 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
311 		return r;
312 	}
313 
314 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
315 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
316 	memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
317 		    adev->vce.fw->size - offset);
318 
319 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
320 
321 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
322 
323 	return 0;
324 }
325 
326 /**
327  * amdgpu_vce_idle_work_handler - power off VCE
328  *
329  * @work: pointer to work structure
330  *
331  * power of VCE when it's not used any more
332  */
333 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
334 {
335 	struct amdgpu_device *adev =
336 		container_of(work, struct amdgpu_device, vce.idle_work.work);
337 	unsigned i, count = 0;
338 
339 	for (i = 0; i < adev->vce.num_rings; i++)
340 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
341 
342 	if (count == 0) {
343 		if (adev->pm.dpm_enabled) {
344 			amdgpu_dpm_enable_vce(adev, false);
345 		} else {
346 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
347 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
348 							       AMD_PG_STATE_GATE);
349 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
350 							       AMD_CG_STATE_GATE);
351 		}
352 	} else {
353 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
354 	}
355 }
356 
357 /**
358  * amdgpu_vce_ring_begin_use - power up VCE
359  *
360  * @ring: amdgpu ring
361  *
362  * Make sure VCE is powerd up when we want to use it
363  */
364 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
365 {
366 	struct amdgpu_device *adev = ring->adev;
367 	bool set_clocks;
368 
369 	if (amdgpu_sriov_vf(adev))
370 		return;
371 
372 	mutex_lock(&adev->vce.idle_mutex);
373 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
374 	if (set_clocks) {
375 		if (adev->pm.dpm_enabled) {
376 			amdgpu_dpm_enable_vce(adev, true);
377 		} else {
378 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
379 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
380 							       AMD_CG_STATE_UNGATE);
381 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
382 							       AMD_PG_STATE_UNGATE);
383 
384 		}
385 	}
386 	mutex_unlock(&adev->vce.idle_mutex);
387 }
388 
389 /**
390  * amdgpu_vce_ring_end_use - power VCE down
391  *
392  * @ring: amdgpu ring
393  *
394  * Schedule work to power VCE down again
395  */
396 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
397 {
398 	if (!amdgpu_sriov_vf(ring->adev))
399 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
400 }
401 
402 /**
403  * amdgpu_vce_free_handles - free still open VCE handles
404  *
405  * @adev: amdgpu_device pointer
406  * @filp: drm file pointer
407  *
408  * Close all VCE handles still open by this file pointer
409  */
410 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
411 {
412 	struct amdgpu_ring *ring = &adev->vce.ring[0];
413 	int i, r;
414 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
415 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
416 
417 		if (!handle || adev->vce.filp[i] != filp)
418 			continue;
419 
420 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
421 		if (r)
422 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
423 
424 		adev->vce.filp[i] = NULL;
425 		atomic_set(&adev->vce.handles[i], 0);
426 	}
427 }
428 
429 /**
430  * amdgpu_vce_get_create_msg - generate a VCE create msg
431  *
432  * @ring: ring we should submit the msg to
433  * @handle: VCE session handle to use
434  * @bo: amdgpu object for which we query the offset
435  * @fence: optional fence to return
436  *
437  * Open up a stream for HW test
438  */
439 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
440 				     struct amdgpu_bo *bo,
441 				     struct dma_fence **fence)
442 {
443 	const unsigned ib_size_dw = 1024;
444 	struct amdgpu_job *job;
445 	struct amdgpu_ib *ib;
446 	struct dma_fence *f = NULL;
447 	uint64_t addr;
448 	int i, r;
449 
450 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4,
451 				     AMDGPU_IB_POOL_DIRECT, &job);
452 	if (r)
453 		return r;
454 
455 	ib = &job->ibs[0];
456 
457 	addr = amdgpu_bo_gpu_offset(bo);
458 
459 	/* stitch together an VCE create msg */
460 	ib->length_dw = 0;
461 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
462 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
463 	ib->ptr[ib->length_dw++] = handle;
464 
465 	if ((ring->adev->vce.fw_version >> 24) >= 52)
466 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
467 	else
468 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
469 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
470 	ib->ptr[ib->length_dw++] = 0x00000000;
471 	ib->ptr[ib->length_dw++] = 0x00000042;
472 	ib->ptr[ib->length_dw++] = 0x0000000a;
473 	ib->ptr[ib->length_dw++] = 0x00000001;
474 	ib->ptr[ib->length_dw++] = 0x00000080;
475 	ib->ptr[ib->length_dw++] = 0x00000060;
476 	ib->ptr[ib->length_dw++] = 0x00000100;
477 	ib->ptr[ib->length_dw++] = 0x00000100;
478 	ib->ptr[ib->length_dw++] = 0x0000000c;
479 	ib->ptr[ib->length_dw++] = 0x00000000;
480 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
481 		ib->ptr[ib->length_dw++] = 0x00000000;
482 		ib->ptr[ib->length_dw++] = 0x00000000;
483 		ib->ptr[ib->length_dw++] = 0x00000000;
484 		ib->ptr[ib->length_dw++] = 0x00000000;
485 	}
486 
487 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
488 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
489 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
490 	ib->ptr[ib->length_dw++] = addr;
491 	ib->ptr[ib->length_dw++] = 0x00000001;
492 
493 	for (i = ib->length_dw; i < ib_size_dw; ++i)
494 		ib->ptr[i] = 0x0;
495 
496 	r = amdgpu_job_submit_direct(job, ring, &f);
497 	if (r)
498 		goto err;
499 
500 	if (fence)
501 		*fence = dma_fence_get(f);
502 	dma_fence_put(f);
503 	return 0;
504 
505 err:
506 	amdgpu_job_free(job);
507 	return r;
508 }
509 
510 /**
511  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
512  *
513  * @ring: ring we should submit the msg to
514  * @handle: VCE session handle to use
515  * @direct: direct or delayed pool
516  * @fence: optional fence to return
517  *
518  * Close up a stream for HW test or if userspace failed to do so
519  */
520 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
521 				      bool direct, struct dma_fence **fence)
522 {
523 	const unsigned ib_size_dw = 1024;
524 	struct amdgpu_job *job;
525 	struct amdgpu_ib *ib;
526 	struct dma_fence *f = NULL;
527 	int i, r;
528 
529 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4,
530 				     direct ? AMDGPU_IB_POOL_DIRECT :
531 				     AMDGPU_IB_POOL_DELAYED, &job);
532 	if (r)
533 		return r;
534 
535 	ib = &job->ibs[0];
536 
537 	/* stitch together an VCE destroy msg */
538 	ib->length_dw = 0;
539 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
540 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
541 	ib->ptr[ib->length_dw++] = handle;
542 
543 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
544 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
545 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
546 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
547 	ib->ptr[ib->length_dw++] = 0x00000000;
548 	ib->ptr[ib->length_dw++] = 0x00000000;
549 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
550 	ib->ptr[ib->length_dw++] = 0x00000000;
551 
552 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
553 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
554 
555 	for (i = ib->length_dw; i < ib_size_dw; ++i)
556 		ib->ptr[i] = 0x0;
557 
558 	if (direct)
559 		r = amdgpu_job_submit_direct(job, ring, &f);
560 	else
561 		r = amdgpu_job_submit(job, &ring->adev->vce.entity,
562 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
563 	if (r)
564 		goto err;
565 
566 	if (fence)
567 		*fence = dma_fence_get(f);
568 	dma_fence_put(f);
569 	return 0;
570 
571 err:
572 	amdgpu_job_free(job);
573 	return r;
574 }
575 
576 /**
577  * amdgpu_vce_cs_validate_bo - make sure not to cross 4GB boundary
578  *
579  * @p: parser context
580  * @ib_idx: indirect buffer to use
581  * @lo: address of lower dword
582  * @hi: address of higher dword
583  * @size: minimum size
584  * @index: bs/fb index
585  *
586  * Make sure that no BO cross a 4GB boundary.
587  */
588 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p, uint32_t ib_idx,
589 				  int lo, int hi, unsigned size, int32_t index)
590 {
591 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
592 	struct ttm_operation_ctx ctx = { false, false };
593 	struct amdgpu_bo_va_mapping *mapping;
594 	unsigned i, fpfn, lpfn;
595 	struct amdgpu_bo *bo;
596 	uint64_t addr;
597 	int r;
598 
599 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
600 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
601 	if (index >= 0) {
602 		addr += offset;
603 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
604 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
605 	} else {
606 		fpfn = 0;
607 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
608 	}
609 
610 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
611 	if (r) {
612 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
613 			  addr, lo, hi, size, index);
614 		return r;
615 	}
616 
617 	for (i = 0; i < bo->placement.num_placement; ++i) {
618 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
619 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
620 			min(bo->placements[i].lpfn, lpfn) : lpfn;
621 	}
622 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
623 }
624 
625 
626 /**
627  * amdgpu_vce_cs_reloc - command submission relocation
628  *
629  * @p: parser context
630  * @ib_idx: indirect buffer to use
631  * @lo: address of lower dword
632  * @hi: address of higher dword
633  * @size: minimum size
634  * @index: bs/fb index
635  *
636  * Patch relocation inside command stream with real buffer address
637  */
638 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
639 			       int lo, int hi, unsigned size, uint32_t index)
640 {
641 	struct amdgpu_bo_va_mapping *mapping;
642 	struct amdgpu_bo *bo;
643 	uint64_t addr;
644 	int r;
645 
646 	if (index == 0xffffffff)
647 		index = 0;
648 
649 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
650 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
651 	addr += ((uint64_t)size) * ((uint64_t)index);
652 
653 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
654 	if (r) {
655 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
656 			  addr, lo, hi, size, index);
657 		return r;
658 	}
659 
660 	if ((addr + (uint64_t)size) >
661 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
662 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
663 			  addr, lo, hi);
664 		return -EINVAL;
665 	}
666 
667 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
668 	addr += amdgpu_bo_gpu_offset(bo);
669 	addr -= ((uint64_t)size) * ((uint64_t)index);
670 
671 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
672 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
673 
674 	return 0;
675 }
676 
677 /**
678  * amdgpu_vce_validate_handle - validate stream handle
679  *
680  * @p: parser context
681  * @handle: handle to validate
682  * @allocated: allocated a new handle?
683  *
684  * Validates the handle and return the found session index or -EINVAL
685  * we we don't have another free session index.
686  */
687 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
688 				      uint32_t handle, uint32_t *allocated)
689 {
690 	unsigned i;
691 
692 	/* validate the handle */
693 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
694 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
695 			if (p->adev->vce.filp[i] != p->filp) {
696 				DRM_ERROR("VCE handle collision detected!\n");
697 				return -EINVAL;
698 			}
699 			return i;
700 		}
701 	}
702 
703 	/* handle not found try to alloc a new one */
704 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
705 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
706 			p->adev->vce.filp[i] = p->filp;
707 			p->adev->vce.img_size[i] = 0;
708 			*allocated |= 1 << i;
709 			return i;
710 		}
711 	}
712 
713 	DRM_ERROR("No more free VCE handles!\n");
714 	return -EINVAL;
715 }
716 
717 /**
718  * amdgpu_vce_cs_parse - parse and validate the command stream
719  *
720  * @p: parser context
721  * @ib_idx: indirect buffer to use
722  */
723 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
724 {
725 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
726 	unsigned fb_idx = 0, bs_idx = 0;
727 	int session_idx = -1;
728 	uint32_t destroyed = 0;
729 	uint32_t created = 0;
730 	uint32_t allocated = 0;
731 	uint32_t tmp, handle = 0;
732 	uint32_t *size = &tmp;
733 	unsigned idx;
734 	int i, r = 0;
735 
736 	p->job->vm = NULL;
737 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
738 
739 	for (idx = 0; idx < ib->length_dw;) {
740 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
741 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
742 
743 		if ((len < 8) || (len & 3)) {
744 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
745 			r = -EINVAL;
746 			goto out;
747 		}
748 
749 		switch (cmd) {
750 		case 0x00000002: /* task info */
751 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
752 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
753 			break;
754 
755 		case 0x03000001: /* encode */
756 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 10,
757 						   idx + 9, 0, 0);
758 			if (r)
759 				goto out;
760 
761 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 12,
762 						   idx + 11, 0, 0);
763 			if (r)
764 				goto out;
765 			break;
766 
767 		case 0x05000001: /* context buffer */
768 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
769 						   idx + 2, 0, 0);
770 			if (r)
771 				goto out;
772 			break;
773 
774 		case 0x05000004: /* video bitstream buffer */
775 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
776 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
777 						   tmp, bs_idx);
778 			if (r)
779 				goto out;
780 			break;
781 
782 		case 0x05000005: /* feedback buffer */
783 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
784 						   4096, fb_idx);
785 			if (r)
786 				goto out;
787 			break;
788 
789 		case 0x0500000d: /* MV buffer */
790 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
791 							idx + 2, 0, 0);
792 			if (r)
793 				goto out;
794 
795 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 8,
796 							idx + 7, 0, 0);
797 			if (r)
798 				goto out;
799 			break;
800 		}
801 
802 		idx += len / 4;
803 	}
804 
805 	for (idx = 0; idx < ib->length_dw;) {
806 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
807 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
808 
809 		switch (cmd) {
810 		case 0x00000001: /* session */
811 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
812 			session_idx = amdgpu_vce_validate_handle(p, handle,
813 								 &allocated);
814 			if (session_idx < 0) {
815 				r = session_idx;
816 				goto out;
817 			}
818 			size = &p->adev->vce.img_size[session_idx];
819 			break;
820 
821 		case 0x00000002: /* task info */
822 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
823 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
824 			break;
825 
826 		case 0x01000001: /* create */
827 			created |= 1 << session_idx;
828 			if (destroyed & (1 << session_idx)) {
829 				destroyed &= ~(1 << session_idx);
830 				allocated |= 1 << session_idx;
831 
832 			} else if (!(allocated & (1 << session_idx))) {
833 				DRM_ERROR("Handle already in use!\n");
834 				r = -EINVAL;
835 				goto out;
836 			}
837 
838 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
839 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
840 				8 * 3 / 2;
841 			break;
842 
843 		case 0x04000001: /* config extension */
844 		case 0x04000002: /* pic control */
845 		case 0x04000005: /* rate control */
846 		case 0x04000007: /* motion estimation */
847 		case 0x04000008: /* rdo */
848 		case 0x04000009: /* vui */
849 		case 0x05000002: /* auxiliary buffer */
850 		case 0x05000009: /* clock table */
851 			break;
852 
853 		case 0x0500000c: /* hw config */
854 			switch (p->adev->asic_type) {
855 #ifdef CONFIG_DRM_AMDGPU_CIK
856 			case CHIP_KAVERI:
857 			case CHIP_MULLINS:
858 #endif
859 			case CHIP_CARRIZO:
860 				break;
861 			default:
862 				r = -EINVAL;
863 				goto out;
864 			}
865 			break;
866 
867 		case 0x03000001: /* encode */
868 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
869 						*size, 0);
870 			if (r)
871 				goto out;
872 
873 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
874 						*size / 3, 0);
875 			if (r)
876 				goto out;
877 			break;
878 
879 		case 0x02000001: /* destroy */
880 			destroyed |= 1 << session_idx;
881 			break;
882 
883 		case 0x05000001: /* context buffer */
884 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
885 						*size * 2, 0);
886 			if (r)
887 				goto out;
888 			break;
889 
890 		case 0x05000004: /* video bitstream buffer */
891 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
892 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
893 						tmp, bs_idx);
894 			if (r)
895 				goto out;
896 			break;
897 
898 		case 0x05000005: /* feedback buffer */
899 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
900 						4096, fb_idx);
901 			if (r)
902 				goto out;
903 			break;
904 
905 		case 0x0500000d: /* MV buffer */
906 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3,
907 							idx + 2, *size, 0);
908 			if (r)
909 				goto out;
910 
911 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 8,
912 							idx + 7, *size / 12, 0);
913 			if (r)
914 				goto out;
915 			break;
916 
917 		default:
918 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
919 			r = -EINVAL;
920 			goto out;
921 		}
922 
923 		if (session_idx == -1) {
924 			DRM_ERROR("no session command at start of IB\n");
925 			r = -EINVAL;
926 			goto out;
927 		}
928 
929 		idx += len / 4;
930 	}
931 
932 	if (allocated & ~created) {
933 		DRM_ERROR("New session without create command!\n");
934 		r = -ENOENT;
935 	}
936 
937 out:
938 	if (!r) {
939 		/* No error, free all destroyed handle slots */
940 		tmp = destroyed;
941 	} else {
942 		/* Error during parsing, free all allocated handle slots */
943 		tmp = allocated;
944 	}
945 
946 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
947 		if (tmp & (1 << i))
948 			atomic_set(&p->adev->vce.handles[i], 0);
949 
950 	return r;
951 }
952 
953 /**
954  * amdgpu_vce_cs_parse_vm - parse the command stream in VM mode
955  *
956  * @p: parser context
957  * @ib_idx: indirect buffer to use
958  */
959 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p, uint32_t ib_idx)
960 {
961 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
962 	int session_idx = -1;
963 	uint32_t destroyed = 0;
964 	uint32_t created = 0;
965 	uint32_t allocated = 0;
966 	uint32_t tmp, handle = 0;
967 	int i, r = 0, idx = 0;
968 
969 	while (idx < ib->length_dw) {
970 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
971 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
972 
973 		if ((len < 8) || (len & 3)) {
974 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
975 			r = -EINVAL;
976 			goto out;
977 		}
978 
979 		switch (cmd) {
980 		case 0x00000001: /* session */
981 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
982 			session_idx = amdgpu_vce_validate_handle(p, handle,
983 								 &allocated);
984 			if (session_idx < 0) {
985 				r = session_idx;
986 				goto out;
987 			}
988 			break;
989 
990 		case 0x01000001: /* create */
991 			created |= 1 << session_idx;
992 			if (destroyed & (1 << session_idx)) {
993 				destroyed &= ~(1 << session_idx);
994 				allocated |= 1 << session_idx;
995 
996 			} else if (!(allocated & (1 << session_idx))) {
997 				DRM_ERROR("Handle already in use!\n");
998 				r = -EINVAL;
999 				goto out;
1000 			}
1001 
1002 			break;
1003 
1004 		case 0x02000001: /* destroy */
1005 			destroyed |= 1 << session_idx;
1006 			break;
1007 
1008 		default:
1009 			break;
1010 		}
1011 
1012 		if (session_idx == -1) {
1013 			DRM_ERROR("no session command at start of IB\n");
1014 			r = -EINVAL;
1015 			goto out;
1016 		}
1017 
1018 		idx += len / 4;
1019 	}
1020 
1021 	if (allocated & ~created) {
1022 		DRM_ERROR("New session without create command!\n");
1023 		r = -ENOENT;
1024 	}
1025 
1026 out:
1027 	if (!r) {
1028 		/* No error, free all destroyed handle slots */
1029 		tmp = destroyed;
1030 		amdgpu_ib_free(p->adev, ib, NULL);
1031 	} else {
1032 		/* Error during parsing, free all allocated handle slots */
1033 		tmp = allocated;
1034 	}
1035 
1036 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1037 		if (tmp & (1 << i))
1038 			atomic_set(&p->adev->vce.handles[i], 0);
1039 
1040 	return r;
1041 }
1042 
1043 /**
1044  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1045  *
1046  * @ring: engine to use
1047  * @job: job to retrieve vmid from
1048  * @ib: the IB to execute
1049  * @flags: unused
1050  *
1051  */
1052 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1053 				struct amdgpu_job *job,
1054 				struct amdgpu_ib *ib,
1055 				uint32_t flags)
1056 {
1057 	amdgpu_ring_write(ring, VCE_CMD_IB);
1058 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1059 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1060 	amdgpu_ring_write(ring, ib->length_dw);
1061 }
1062 
1063 /**
1064  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1065  *
1066  * @ring: engine to use
1067  * @addr: address
1068  * @seq: sequence number
1069  * @flags: fence related flags
1070  *
1071  */
1072 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1073 				unsigned flags)
1074 {
1075 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1076 
1077 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1078 	amdgpu_ring_write(ring, addr);
1079 	amdgpu_ring_write(ring, upper_32_bits(addr));
1080 	amdgpu_ring_write(ring, seq);
1081 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1082 	amdgpu_ring_write(ring, VCE_CMD_END);
1083 }
1084 
1085 /**
1086  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1087  *
1088  * @ring: the engine to test on
1089  *
1090  */
1091 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1092 {
1093 	struct amdgpu_device *adev = ring->adev;
1094 	uint32_t rptr;
1095 	unsigned i;
1096 	int r, timeout = adev->usec_timeout;
1097 
1098 	/* skip ring test for sriov*/
1099 	if (amdgpu_sriov_vf(adev))
1100 		return 0;
1101 
1102 	r = amdgpu_ring_alloc(ring, 16);
1103 	if (r)
1104 		return r;
1105 
1106 	rptr = amdgpu_ring_get_rptr(ring);
1107 
1108 	amdgpu_ring_write(ring, VCE_CMD_END);
1109 	amdgpu_ring_commit(ring);
1110 
1111 	for (i = 0; i < timeout; i++) {
1112 		if (amdgpu_ring_get_rptr(ring) != rptr)
1113 			break;
1114 		udelay(1);
1115 	}
1116 
1117 	if (i >= timeout)
1118 		r = -ETIMEDOUT;
1119 
1120 	return r;
1121 }
1122 
1123 /**
1124  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1125  *
1126  * @ring: the engine to test on
1127  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1128  *
1129  */
1130 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1131 {
1132 	struct dma_fence *fence = NULL;
1133 	struct amdgpu_bo *bo = NULL;
1134 	long r;
1135 
1136 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1137 	if (ring != &ring->adev->vce.ring[0])
1138 		return 0;
1139 
1140 	r = amdgpu_bo_create_reserved(ring->adev, 512, PAGE_SIZE,
1141 				      AMDGPU_GEM_DOMAIN_VRAM,
1142 				      &bo, NULL, NULL);
1143 	if (r)
1144 		return r;
1145 
1146 	r = amdgpu_vce_get_create_msg(ring, 1, bo, NULL);
1147 	if (r)
1148 		goto error;
1149 
1150 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1151 	if (r)
1152 		goto error;
1153 
1154 	r = dma_fence_wait_timeout(fence, false, timeout);
1155 	if (r == 0)
1156 		r = -ETIMEDOUT;
1157 	else if (r > 0)
1158 		r = 0;
1159 
1160 error:
1161 	dma_fence_put(fence);
1162 	amdgpu_bo_unreserve(bo);
1163 	amdgpu_bo_free_kernel(&bo, NULL, NULL);
1164 	return r;
1165 }
1166