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