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 #include <drm/drm_drv.h>
33 
34 #include "amdgpu.h"
35 #include "amdgpu_pm.h"
36 #include "amdgpu_vce.h"
37 #include "amdgpu_cs.h"
38 #include "cikd.h"
39 
40 /* 1 second timeout */
41 #define VCE_IDLE_TIMEOUT	msecs_to_jiffies(1000)
42 
43 /* Firmware Names */
44 #ifdef CONFIG_DRM_AMDGPU_CIK
45 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_vce.bin"
46 #define FIRMWARE_KABINI	"amdgpu/kabini_vce.bin"
47 #define FIRMWARE_KAVERI	"amdgpu/kaveri_vce.bin"
48 #define FIRMWARE_HAWAII	"amdgpu/hawaii_vce.bin"
49 #define FIRMWARE_MULLINS	"amdgpu/mullins_vce.bin"
50 #endif
51 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
52 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
53 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
54 #define FIRMWARE_STONEY		"amdgpu/stoney_vce.bin"
55 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_vce.bin"
56 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_vce.bin"
57 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_vce.bin"
58 #define FIRMWARE_VEGAM		"amdgpu/vegam_vce.bin"
59 
60 #define FIRMWARE_VEGA10		"amdgpu/vega10_vce.bin"
61 #define FIRMWARE_VEGA12		"amdgpu/vega12_vce.bin"
62 #define FIRMWARE_VEGA20		"amdgpu/vega20_vce.bin"
63 
64 #ifdef CONFIG_DRM_AMDGPU_CIK
65 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
66 MODULE_FIRMWARE(FIRMWARE_KABINI);
67 MODULE_FIRMWARE(FIRMWARE_KAVERI);
68 MODULE_FIRMWARE(FIRMWARE_HAWAII);
69 MODULE_FIRMWARE(FIRMWARE_MULLINS);
70 #endif
71 MODULE_FIRMWARE(FIRMWARE_TONGA);
72 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
73 MODULE_FIRMWARE(FIRMWARE_FIJI);
74 MODULE_FIRMWARE(FIRMWARE_STONEY);
75 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
76 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
77 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
78 MODULE_FIRMWARE(FIRMWARE_VEGAM);
79 
80 MODULE_FIRMWARE(FIRMWARE_VEGA10);
81 MODULE_FIRMWARE(FIRMWARE_VEGA12);
82 MODULE_FIRMWARE(FIRMWARE_VEGA20);
83 
84 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
85 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
86 				     struct dma_fence **fence);
87 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
88 				      bool direct, struct dma_fence **fence);
89 
90 /**
91  * amdgpu_vce_sw_init - allocate memory, load vce firmware
92  *
93  * @adev: amdgpu_device pointer
94  * @size: size for the new BO
95  *
96  * First step to get VCE online, allocate memory and load the firmware
97  */
98 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
99 {
100 	const char *fw_name;
101 	const struct common_firmware_header *hdr;
102 	unsigned ucode_version, version_major, version_minor, binary_id;
103 	int i, r;
104 
105 	switch (adev->asic_type) {
106 #ifdef CONFIG_DRM_AMDGPU_CIK
107 	case CHIP_BONAIRE:
108 		fw_name = FIRMWARE_BONAIRE;
109 		break;
110 	case CHIP_KAVERI:
111 		fw_name = FIRMWARE_KAVERI;
112 		break;
113 	case CHIP_KABINI:
114 		fw_name = FIRMWARE_KABINI;
115 		break;
116 	case CHIP_HAWAII:
117 		fw_name = FIRMWARE_HAWAII;
118 		break;
119 	case CHIP_MULLINS:
120 		fw_name = FIRMWARE_MULLINS;
121 		break;
122 #endif
123 	case CHIP_TONGA:
124 		fw_name = FIRMWARE_TONGA;
125 		break;
126 	case CHIP_CARRIZO:
127 		fw_name = FIRMWARE_CARRIZO;
128 		break;
129 	case CHIP_FIJI:
130 		fw_name = FIRMWARE_FIJI;
131 		break;
132 	case CHIP_STONEY:
133 		fw_name = FIRMWARE_STONEY;
134 		break;
135 	case CHIP_POLARIS10:
136 		fw_name = FIRMWARE_POLARIS10;
137 		break;
138 	case CHIP_POLARIS11:
139 		fw_name = FIRMWARE_POLARIS11;
140 		break;
141 	case CHIP_POLARIS12:
142 		fw_name = FIRMWARE_POLARIS12;
143 		break;
144 	case CHIP_VEGAM:
145 		fw_name = FIRMWARE_VEGAM;
146 		break;
147 	case CHIP_VEGA10:
148 		fw_name = FIRMWARE_VEGA10;
149 		break;
150 	case CHIP_VEGA12:
151 		fw_name = FIRMWARE_VEGA12;
152 		break;
153 	case CHIP_VEGA20:
154 		fw_name = FIRMWARE_VEGA20;
155 		break;
156 
157 	default:
158 		return -EINVAL;
159 	}
160 
161 	r = amdgpu_ucode_request(adev, &adev->vce.fw, fw_name);
162 	if (r) {
163 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
164 			fw_name);
165 		amdgpu_ucode_release(&adev->vce.fw);
166 		return r;
167 	}
168 
169 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
170 
171 	ucode_version = le32_to_cpu(hdr->ucode_version);
172 	version_major = (ucode_version >> 20) & 0xfff;
173 	version_minor = (ucode_version >> 8) & 0xfff;
174 	binary_id = ucode_version & 0xff;
175 	DRM_INFO("Found VCE firmware Version: %d.%d Binary ID: %d\n",
176 		version_major, version_minor, binary_id);
177 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
178 				(binary_id << 8));
179 
180 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
181 				    AMDGPU_GEM_DOMAIN_VRAM |
182 				    AMDGPU_GEM_DOMAIN_GTT,
183 				    &adev->vce.vcpu_bo,
184 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
185 	if (r) {
186 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
187 		return r;
188 	}
189 
190 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
191 		atomic_set(&adev->vce.handles[i], 0);
192 		adev->vce.filp[i] = NULL;
193 	}
194 
195 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
196 	mutex_init(&adev->vce.idle_mutex);
197 
198 	return 0;
199 }
200 
201 /**
202  * amdgpu_vce_sw_fini - free memory
203  *
204  * @adev: amdgpu_device pointer
205  *
206  * Last step on VCE teardown, free firmware memory
207  */
208 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
209 {
210 	unsigned i;
211 
212 	if (adev->vce.vcpu_bo == NULL)
213 		return 0;
214 
215 	drm_sched_entity_destroy(&adev->vce.entity);
216 
217 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
218 		(void **)&adev->vce.cpu_addr);
219 
220 	for (i = 0; i < adev->vce.num_rings; i++)
221 		amdgpu_ring_fini(&adev->vce.ring[i]);
222 
223 	amdgpu_ucode_release(&adev->vce.fw);
224 	mutex_destroy(&adev->vce.idle_mutex);
225 
226 	return 0;
227 }
228 
229 /**
230  * amdgpu_vce_entity_init - init entity
231  *
232  * @adev: amdgpu_device pointer
233  *
234  */
235 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
236 {
237 	struct amdgpu_ring *ring;
238 	struct drm_gpu_scheduler *sched;
239 	int r;
240 
241 	ring = &adev->vce.ring[0];
242 	sched = &ring->sched;
243 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
244 				  &sched, 1, NULL);
245 	if (r != 0) {
246 		DRM_ERROR("Failed setting up VCE run queue.\n");
247 		return r;
248 	}
249 
250 	return 0;
251 }
252 
253 /**
254  * amdgpu_vce_suspend - unpin VCE fw memory
255  *
256  * @adev: amdgpu_device pointer
257  *
258  */
259 int amdgpu_vce_suspend(struct amdgpu_device *adev)
260 {
261 	int i;
262 
263 	cancel_delayed_work_sync(&adev->vce.idle_work);
264 
265 	if (adev->vce.vcpu_bo == NULL)
266 		return 0;
267 
268 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
269 		if (atomic_read(&adev->vce.handles[i]))
270 			break;
271 
272 	if (i == AMDGPU_MAX_VCE_HANDLES)
273 		return 0;
274 
275 	/* TODO: suspending running encoding sessions isn't supported */
276 	return -EINVAL;
277 }
278 
279 /**
280  * amdgpu_vce_resume - pin VCE fw memory
281  *
282  * @adev: amdgpu_device pointer
283  *
284  */
285 int amdgpu_vce_resume(struct amdgpu_device *adev)
286 {
287 	void *cpu_addr;
288 	const struct common_firmware_header *hdr;
289 	unsigned offset;
290 	int r, idx;
291 
292 	if (adev->vce.vcpu_bo == NULL)
293 		return -EINVAL;
294 
295 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
296 	if (r) {
297 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
298 		return r;
299 	}
300 
301 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
302 	if (r) {
303 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
304 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
305 		return r;
306 	}
307 
308 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
309 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
310 
311 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
312 		memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
313 			    adev->vce.fw->size - offset);
314 		drm_dev_exit(idx);
315 	}
316 
317 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
318 
319 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
320 
321 	return 0;
322 }
323 
324 /**
325  * amdgpu_vce_idle_work_handler - power off VCE
326  *
327  * @work: pointer to work structure
328  *
329  * power of VCE when it's not used any more
330  */
331 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
332 {
333 	struct amdgpu_device *adev =
334 		container_of(work, struct amdgpu_device, vce.idle_work.work);
335 	unsigned i, count = 0;
336 
337 	for (i = 0; i < adev->vce.num_rings; i++)
338 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
339 
340 	if (count == 0) {
341 		if (adev->pm.dpm_enabled) {
342 			amdgpu_dpm_enable_vce(adev, false);
343 		} else {
344 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
345 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
346 							       AMD_PG_STATE_GATE);
347 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
348 							       AMD_CG_STATE_GATE);
349 		}
350 	} else {
351 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
352 	}
353 }
354 
355 /**
356  * amdgpu_vce_ring_begin_use - power up VCE
357  *
358  * @ring: amdgpu ring
359  *
360  * Make sure VCE is powerd up when we want to use it
361  */
362 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
363 {
364 	struct amdgpu_device *adev = ring->adev;
365 	bool set_clocks;
366 
367 	if (amdgpu_sriov_vf(adev))
368 		return;
369 
370 	mutex_lock(&adev->vce.idle_mutex);
371 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
372 	if (set_clocks) {
373 		if (adev->pm.dpm_enabled) {
374 			amdgpu_dpm_enable_vce(adev, true);
375 		} else {
376 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
377 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
378 							       AMD_CG_STATE_UNGATE);
379 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
380 							       AMD_PG_STATE_UNGATE);
381 
382 		}
383 	}
384 	mutex_unlock(&adev->vce.idle_mutex);
385 }
386 
387 /**
388  * amdgpu_vce_ring_end_use - power VCE down
389  *
390  * @ring: amdgpu ring
391  *
392  * Schedule work to power VCE down again
393  */
394 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
395 {
396 	if (!amdgpu_sriov_vf(ring->adev))
397 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
398 }
399 
400 /**
401  * amdgpu_vce_free_handles - free still open VCE handles
402  *
403  * @adev: amdgpu_device pointer
404  * @filp: drm file pointer
405  *
406  * Close all VCE handles still open by this file pointer
407  */
408 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
409 {
410 	struct amdgpu_ring *ring = &adev->vce.ring[0];
411 	int i, r;
412 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
413 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
414 
415 		if (!handle || adev->vce.filp[i] != filp)
416 			continue;
417 
418 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
419 		if (r)
420 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
421 
422 		adev->vce.filp[i] = NULL;
423 		atomic_set(&adev->vce.handles[i], 0);
424 	}
425 }
426 
427 /**
428  * amdgpu_vce_get_create_msg - generate a VCE create msg
429  *
430  * @ring: ring we should submit the msg to
431  * @handle: VCE session handle to use
432  * @fence: optional fence to return
433  *
434  * Open up a stream for HW test
435  */
436 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
437 				     struct dma_fence **fence)
438 {
439 	const unsigned ib_size_dw = 1024;
440 	struct amdgpu_job *job;
441 	struct amdgpu_ib *ib;
442 	struct amdgpu_ib ib_msg;
443 	struct dma_fence *f = NULL;
444 	uint64_t addr;
445 	int i, r;
446 
447 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
448 				     AMDGPU_FENCE_OWNER_UNDEFINED,
449 				     ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT,
450 				     &job);
451 	if (r)
452 		return r;
453 
454 	memset(&ib_msg, 0, sizeof(ib_msg));
455 	/* only one gpu page is needed, alloc +1 page to make addr aligned. */
456 	r = amdgpu_ib_get(ring->adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2,
457 			  AMDGPU_IB_POOL_DIRECT,
458 			  &ib_msg);
459 	if (r)
460 		goto err;
461 
462 	ib = &job->ibs[0];
463 	/* let addr point to page boundary */
464 	addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg.gpu_addr);
465 
466 	/* stitch together an VCE create msg */
467 	ib->length_dw = 0;
468 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
469 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
470 	ib->ptr[ib->length_dw++] = handle;
471 
472 	if ((ring->adev->vce.fw_version >> 24) >= 52)
473 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
474 	else
475 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
476 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
477 	ib->ptr[ib->length_dw++] = 0x00000000;
478 	ib->ptr[ib->length_dw++] = 0x00000042;
479 	ib->ptr[ib->length_dw++] = 0x0000000a;
480 	ib->ptr[ib->length_dw++] = 0x00000001;
481 	ib->ptr[ib->length_dw++] = 0x00000080;
482 	ib->ptr[ib->length_dw++] = 0x00000060;
483 	ib->ptr[ib->length_dw++] = 0x00000100;
484 	ib->ptr[ib->length_dw++] = 0x00000100;
485 	ib->ptr[ib->length_dw++] = 0x0000000c;
486 	ib->ptr[ib->length_dw++] = 0x00000000;
487 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
488 		ib->ptr[ib->length_dw++] = 0x00000000;
489 		ib->ptr[ib->length_dw++] = 0x00000000;
490 		ib->ptr[ib->length_dw++] = 0x00000000;
491 		ib->ptr[ib->length_dw++] = 0x00000000;
492 	}
493 
494 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
495 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
496 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
497 	ib->ptr[ib->length_dw++] = addr;
498 	ib->ptr[ib->length_dw++] = 0x00000001;
499 
500 	for (i = ib->length_dw; i < ib_size_dw; ++i)
501 		ib->ptr[i] = 0x0;
502 
503 	r = amdgpu_job_submit_direct(job, ring, &f);
504 	amdgpu_ib_free(ring->adev, &ib_msg, f);
505 	if (r)
506 		goto err;
507 
508 	if (fence)
509 		*fence = dma_fence_get(f);
510 	dma_fence_put(f);
511 	return 0;
512 
513 err:
514 	amdgpu_job_free(job);
515 	return r;
516 }
517 
518 /**
519  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
520  *
521  * @ring: ring we should submit the msg to
522  * @handle: VCE session handle to use
523  * @direct: direct or delayed pool
524  * @fence: optional fence to return
525  *
526  * Close up a stream for HW test or if userspace failed to do so
527  */
528 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
529 				      bool direct, struct dma_fence **fence)
530 {
531 	const unsigned ib_size_dw = 1024;
532 	struct amdgpu_job *job;
533 	struct amdgpu_ib *ib;
534 	struct dma_fence *f = NULL;
535 	int i, r;
536 
537 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
538 				     AMDGPU_FENCE_OWNER_UNDEFINED,
539 				     ib_size_dw * 4,
540 				     direct ? AMDGPU_IB_POOL_DIRECT :
541 				     AMDGPU_IB_POOL_DELAYED, &job);
542 	if (r)
543 		return r;
544 
545 	ib = &job->ibs[0];
546 
547 	/* stitch together an VCE destroy msg */
548 	ib->length_dw = 0;
549 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
550 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
551 	ib->ptr[ib->length_dw++] = handle;
552 
553 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
554 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
555 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
556 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
557 	ib->ptr[ib->length_dw++] = 0x00000000;
558 	ib->ptr[ib->length_dw++] = 0x00000000;
559 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
560 	ib->ptr[ib->length_dw++] = 0x00000000;
561 
562 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
563 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
564 
565 	for (i = ib->length_dw; i < ib_size_dw; ++i)
566 		ib->ptr[i] = 0x0;
567 
568 	if (direct)
569 		r = amdgpu_job_submit_direct(job, ring, &f);
570 	else
571 		f = amdgpu_job_submit(job);
572 	if (r)
573 		goto err;
574 
575 	if (fence)
576 		*fence = dma_fence_get(f);
577 	dma_fence_put(f);
578 	return 0;
579 
580 err:
581 	amdgpu_job_free(job);
582 	return r;
583 }
584 
585 /**
586  * amdgpu_vce_validate_bo - make sure not to cross 4GB boundary
587  *
588  * @ib: indirect buffer to use
589  * @lo: address of lower dword
590  * @hi: address of higher dword
591  * @size: minimum size
592  * @index: bs/fb index
593  *
594  * Make sure that no BO cross a 4GB boundary.
595  */
596 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p,
597 				  struct amdgpu_ib *ib, int lo, int hi,
598 				  unsigned size, int32_t index)
599 {
600 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
601 	struct ttm_operation_ctx ctx = { false, false };
602 	struct amdgpu_bo_va_mapping *mapping;
603 	unsigned i, fpfn, lpfn;
604 	struct amdgpu_bo *bo;
605 	uint64_t addr;
606 	int r;
607 
608 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
609 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
610 	if (index >= 0) {
611 		addr += offset;
612 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
613 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
614 	} else {
615 		fpfn = 0;
616 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
617 	}
618 
619 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
620 	if (r) {
621 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
622 			  addr, lo, hi, size, index);
623 		return r;
624 	}
625 
626 	for (i = 0; i < bo->placement.num_placement; ++i) {
627 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
628 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
629 			min(bo->placements[i].lpfn, lpfn) : lpfn;
630 	}
631 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
632 }
633 
634 
635 /**
636  * amdgpu_vce_cs_reloc - command submission relocation
637  *
638  * @p: parser context
639  * @ib: indirect buffer to use
640  * @lo: address of lower dword
641  * @hi: address of higher dword
642  * @size: minimum size
643  * @index: bs/fb index
644  *
645  * Patch relocation inside command stream with real buffer address
646  */
647 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, struct amdgpu_ib *ib,
648 			       int lo, int hi, unsigned size, uint32_t index)
649 {
650 	struct amdgpu_bo_va_mapping *mapping;
651 	struct amdgpu_bo *bo;
652 	uint64_t addr;
653 	int r;
654 
655 	if (index == 0xffffffff)
656 		index = 0;
657 
658 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
659 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
660 	addr += ((uint64_t)size) * ((uint64_t)index);
661 
662 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
663 	if (r) {
664 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
665 			  addr, lo, hi, size, index);
666 		return r;
667 	}
668 
669 	if ((addr + (uint64_t)size) >
670 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
671 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
672 			  addr, lo, hi);
673 		return -EINVAL;
674 	}
675 
676 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
677 	addr += amdgpu_bo_gpu_offset(bo);
678 	addr -= ((uint64_t)size) * ((uint64_t)index);
679 
680 	amdgpu_ib_set_value(ib, lo, lower_32_bits(addr));
681 	amdgpu_ib_set_value(ib, hi, upper_32_bits(addr));
682 
683 	return 0;
684 }
685 
686 /**
687  * amdgpu_vce_validate_handle - validate stream handle
688  *
689  * @p: parser context
690  * @handle: handle to validate
691  * @allocated: allocated a new handle?
692  *
693  * Validates the handle and return the found session index or -EINVAL
694  * we we don't have another free session index.
695  */
696 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
697 				      uint32_t handle, uint32_t *allocated)
698 {
699 	unsigned i;
700 
701 	/* validate the handle */
702 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
703 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
704 			if (p->adev->vce.filp[i] != p->filp) {
705 				DRM_ERROR("VCE handle collision detected!\n");
706 				return -EINVAL;
707 			}
708 			return i;
709 		}
710 	}
711 
712 	/* handle not found try to alloc a new one */
713 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
714 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
715 			p->adev->vce.filp[i] = p->filp;
716 			p->adev->vce.img_size[i] = 0;
717 			*allocated |= 1 << i;
718 			return i;
719 		}
720 	}
721 
722 	DRM_ERROR("No more free VCE handles!\n");
723 	return -EINVAL;
724 }
725 
726 /**
727  * amdgpu_vce_ring_parse_cs - parse and validate the command stream
728  *
729  * @p: parser context
730  * @job: the job to parse
731  * @ib: the IB to patch
732  */
733 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p,
734 			     struct amdgpu_job *job,
735 			     struct amdgpu_ib *ib)
736 {
737 	unsigned fb_idx = 0, bs_idx = 0;
738 	int session_idx = -1;
739 	uint32_t destroyed = 0;
740 	uint32_t created = 0;
741 	uint32_t allocated = 0;
742 	uint32_t tmp, handle = 0;
743 	uint32_t *size = &tmp;
744 	unsigned idx;
745 	int i, r = 0;
746 
747 	job->vm = NULL;
748 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
749 
750 	for (idx = 0; idx < ib->length_dw;) {
751 		uint32_t len = amdgpu_ib_get_value(ib, idx);
752 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
753 
754 		if ((len < 8) || (len & 3)) {
755 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
756 			r = -EINVAL;
757 			goto out;
758 		}
759 
760 		switch (cmd) {
761 		case 0x00000002: /* task info */
762 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
763 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
764 			break;
765 
766 		case 0x03000001: /* encode */
767 			r = amdgpu_vce_validate_bo(p, ib, idx + 10, idx + 9,
768 						   0, 0);
769 			if (r)
770 				goto out;
771 
772 			r = amdgpu_vce_validate_bo(p, ib, idx + 12, idx + 11,
773 						   0, 0);
774 			if (r)
775 				goto out;
776 			break;
777 
778 		case 0x05000001: /* context buffer */
779 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
780 						   0, 0);
781 			if (r)
782 				goto out;
783 			break;
784 
785 		case 0x05000004: /* video bitstream buffer */
786 			tmp = amdgpu_ib_get_value(ib, idx + 4);
787 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
788 						   tmp, bs_idx);
789 			if (r)
790 				goto out;
791 			break;
792 
793 		case 0x05000005: /* feedback buffer */
794 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
795 						   4096, fb_idx);
796 			if (r)
797 				goto out;
798 			break;
799 
800 		case 0x0500000d: /* MV buffer */
801 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
802 						   0, 0);
803 			if (r)
804 				goto out;
805 
806 			r = amdgpu_vce_validate_bo(p, ib, idx + 8, idx + 7,
807 						   0, 0);
808 			if (r)
809 				goto out;
810 			break;
811 		}
812 
813 		idx += len / 4;
814 	}
815 
816 	for (idx = 0; idx < ib->length_dw;) {
817 		uint32_t len = amdgpu_ib_get_value(ib, idx);
818 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
819 
820 		switch (cmd) {
821 		case 0x00000001: /* session */
822 			handle = amdgpu_ib_get_value(ib, idx + 2);
823 			session_idx = amdgpu_vce_validate_handle(p, handle,
824 								 &allocated);
825 			if (session_idx < 0) {
826 				r = session_idx;
827 				goto out;
828 			}
829 			size = &p->adev->vce.img_size[session_idx];
830 			break;
831 
832 		case 0x00000002: /* task info */
833 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
834 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
835 			break;
836 
837 		case 0x01000001: /* create */
838 			created |= 1 << session_idx;
839 			if (destroyed & (1 << session_idx)) {
840 				destroyed &= ~(1 << session_idx);
841 				allocated |= 1 << session_idx;
842 
843 			} else if (!(allocated & (1 << session_idx))) {
844 				DRM_ERROR("Handle already in use!\n");
845 				r = -EINVAL;
846 				goto out;
847 			}
848 
849 			*size = amdgpu_ib_get_value(ib, idx + 8) *
850 				amdgpu_ib_get_value(ib, idx + 10) *
851 				8 * 3 / 2;
852 			break;
853 
854 		case 0x04000001: /* config extension */
855 		case 0x04000002: /* pic control */
856 		case 0x04000005: /* rate control */
857 		case 0x04000007: /* motion estimation */
858 		case 0x04000008: /* rdo */
859 		case 0x04000009: /* vui */
860 		case 0x05000002: /* auxiliary buffer */
861 		case 0x05000009: /* clock table */
862 			break;
863 
864 		case 0x0500000c: /* hw config */
865 			switch (p->adev->asic_type) {
866 #ifdef CONFIG_DRM_AMDGPU_CIK
867 			case CHIP_KAVERI:
868 			case CHIP_MULLINS:
869 #endif
870 			case CHIP_CARRIZO:
871 				break;
872 			default:
873 				r = -EINVAL;
874 				goto out;
875 			}
876 			break;
877 
878 		case 0x03000001: /* encode */
879 			r = amdgpu_vce_cs_reloc(p, ib, idx + 10, idx + 9,
880 						*size, 0);
881 			if (r)
882 				goto out;
883 
884 			r = amdgpu_vce_cs_reloc(p, ib, idx + 12, idx + 11,
885 						*size / 3, 0);
886 			if (r)
887 				goto out;
888 			break;
889 
890 		case 0x02000001: /* destroy */
891 			destroyed |= 1 << session_idx;
892 			break;
893 
894 		case 0x05000001: /* context buffer */
895 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
896 						*size * 2, 0);
897 			if (r)
898 				goto out;
899 			break;
900 
901 		case 0x05000004: /* video bitstream buffer */
902 			tmp = amdgpu_ib_get_value(ib, idx + 4);
903 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
904 						tmp, bs_idx);
905 			if (r)
906 				goto out;
907 			break;
908 
909 		case 0x05000005: /* feedback buffer */
910 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
911 						4096, fb_idx);
912 			if (r)
913 				goto out;
914 			break;
915 
916 		case 0x0500000d: /* MV buffer */
917 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3,
918 						idx + 2, *size, 0);
919 			if (r)
920 				goto out;
921 
922 			r = amdgpu_vce_cs_reloc(p, ib, idx + 8,
923 						idx + 7, *size / 12, 0);
924 			if (r)
925 				goto out;
926 			break;
927 
928 		default:
929 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
930 			r = -EINVAL;
931 			goto out;
932 		}
933 
934 		if (session_idx == -1) {
935 			DRM_ERROR("no session command at start of IB\n");
936 			r = -EINVAL;
937 			goto out;
938 		}
939 
940 		idx += len / 4;
941 	}
942 
943 	if (allocated & ~created) {
944 		DRM_ERROR("New session without create command!\n");
945 		r = -ENOENT;
946 	}
947 
948 out:
949 	if (!r) {
950 		/* No error, free all destroyed handle slots */
951 		tmp = destroyed;
952 	} else {
953 		/* Error during parsing, free all allocated handle slots */
954 		tmp = allocated;
955 	}
956 
957 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
958 		if (tmp & (1 << i))
959 			atomic_set(&p->adev->vce.handles[i], 0);
960 
961 	return r;
962 }
963 
964 /**
965  * amdgpu_vce_ring_parse_cs_vm - parse the command stream in VM mode
966  *
967  * @p: parser context
968  * @job: the job to parse
969  * @ib: the IB to patch
970  */
971 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p,
972 				struct amdgpu_job *job,
973 				struct amdgpu_ib *ib)
974 {
975 	int session_idx = -1;
976 	uint32_t destroyed = 0;
977 	uint32_t created = 0;
978 	uint32_t allocated = 0;
979 	uint32_t tmp, handle = 0;
980 	int i, r = 0, idx = 0;
981 
982 	while (idx < ib->length_dw) {
983 		uint32_t len = amdgpu_ib_get_value(ib, idx);
984 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
985 
986 		if ((len < 8) || (len & 3)) {
987 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
988 			r = -EINVAL;
989 			goto out;
990 		}
991 
992 		switch (cmd) {
993 		case 0x00000001: /* session */
994 			handle = amdgpu_ib_get_value(ib, idx + 2);
995 			session_idx = amdgpu_vce_validate_handle(p, handle,
996 								 &allocated);
997 			if (session_idx < 0) {
998 				r = session_idx;
999 				goto out;
1000 			}
1001 			break;
1002 
1003 		case 0x01000001: /* create */
1004 			created |= 1 << session_idx;
1005 			if (destroyed & (1 << session_idx)) {
1006 				destroyed &= ~(1 << session_idx);
1007 				allocated |= 1 << session_idx;
1008 
1009 			} else if (!(allocated & (1 << session_idx))) {
1010 				DRM_ERROR("Handle already in use!\n");
1011 				r = -EINVAL;
1012 				goto out;
1013 			}
1014 
1015 			break;
1016 
1017 		case 0x02000001: /* destroy */
1018 			destroyed |= 1 << session_idx;
1019 			break;
1020 
1021 		default:
1022 			break;
1023 		}
1024 
1025 		if (session_idx == -1) {
1026 			DRM_ERROR("no session command at start of IB\n");
1027 			r = -EINVAL;
1028 			goto out;
1029 		}
1030 
1031 		idx += len / 4;
1032 	}
1033 
1034 	if (allocated & ~created) {
1035 		DRM_ERROR("New session without create command!\n");
1036 		r = -ENOENT;
1037 	}
1038 
1039 out:
1040 	if (!r) {
1041 		/* No error, free all destroyed handle slots */
1042 		tmp = destroyed;
1043 		amdgpu_ib_free(p->adev, ib, NULL);
1044 	} else {
1045 		/* Error during parsing, free all allocated handle slots */
1046 		tmp = allocated;
1047 	}
1048 
1049 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1050 		if (tmp & (1 << i))
1051 			atomic_set(&p->adev->vce.handles[i], 0);
1052 
1053 	return r;
1054 }
1055 
1056 /**
1057  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1058  *
1059  * @ring: engine to use
1060  * @job: job to retrieve vmid from
1061  * @ib: the IB to execute
1062  * @flags: unused
1063  *
1064  */
1065 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1066 				struct amdgpu_job *job,
1067 				struct amdgpu_ib *ib,
1068 				uint32_t flags)
1069 {
1070 	amdgpu_ring_write(ring, VCE_CMD_IB);
1071 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1072 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1073 	amdgpu_ring_write(ring, ib->length_dw);
1074 }
1075 
1076 /**
1077  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1078  *
1079  * @ring: engine to use
1080  * @addr: address
1081  * @seq: sequence number
1082  * @flags: fence related flags
1083  *
1084  */
1085 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1086 				unsigned flags)
1087 {
1088 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1089 
1090 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1091 	amdgpu_ring_write(ring, addr);
1092 	amdgpu_ring_write(ring, upper_32_bits(addr));
1093 	amdgpu_ring_write(ring, seq);
1094 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1095 	amdgpu_ring_write(ring, VCE_CMD_END);
1096 }
1097 
1098 /**
1099  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1100  *
1101  * @ring: the engine to test on
1102  *
1103  */
1104 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1105 {
1106 	struct amdgpu_device *adev = ring->adev;
1107 	uint32_t rptr;
1108 	unsigned i;
1109 	int r, timeout = adev->usec_timeout;
1110 
1111 	/* skip ring test for sriov*/
1112 	if (amdgpu_sriov_vf(adev))
1113 		return 0;
1114 
1115 	r = amdgpu_ring_alloc(ring, 16);
1116 	if (r)
1117 		return r;
1118 
1119 	rptr = amdgpu_ring_get_rptr(ring);
1120 
1121 	amdgpu_ring_write(ring, VCE_CMD_END);
1122 	amdgpu_ring_commit(ring);
1123 
1124 	for (i = 0; i < timeout; i++) {
1125 		if (amdgpu_ring_get_rptr(ring) != rptr)
1126 			break;
1127 		udelay(1);
1128 	}
1129 
1130 	if (i >= timeout)
1131 		r = -ETIMEDOUT;
1132 
1133 	return r;
1134 }
1135 
1136 /**
1137  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1138  *
1139  * @ring: the engine to test on
1140  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1141  *
1142  */
1143 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1144 {
1145 	struct dma_fence *fence = NULL;
1146 	long r;
1147 
1148 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1149 	if (ring != &ring->adev->vce.ring[0])
1150 		return 0;
1151 
1152 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
1153 	if (r)
1154 		goto error;
1155 
1156 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1157 	if (r)
1158 		goto error;
1159 
1160 	r = dma_fence_wait_timeout(fence, false, timeout);
1161 	if (r == 0)
1162 		r = -ETIMEDOUT;
1163 	else if (r > 0)
1164 		r = 0;
1165 
1166 error:
1167 	dma_fence_put(fence);
1168 	return r;
1169 }
1170 
1171 enum amdgpu_ring_priority_level amdgpu_vce_get_ring_prio(int ring)
1172 {
1173 	switch(ring) {
1174 	case 0:
1175 		return AMDGPU_RING_PRIO_0;
1176 	case 1:
1177 		return AMDGPU_RING_PRIO_1;
1178 	case 2:
1179 		return AMDGPU_RING_PRIO_2;
1180 	default:
1181 		return AMDGPU_RING_PRIO_0;
1182 	}
1183 }
1184