xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c (revision 13b9eb15179de69e3c6f7ed714b0499b0abf4394)
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 = request_firmware(&adev->vce.fw, fw_name, adev->dev);
162 	if (r) {
163 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
164 			fw_name);
165 		return r;
166 	}
167 
168 	r = amdgpu_ucode_validate(adev->vce.fw);
169 	if (r) {
170 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
171 			fw_name);
172 		release_firmware(adev->vce.fw);
173 		adev->vce.fw = NULL;
174 		return r;
175 	}
176 
177 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
178 
179 	ucode_version = le32_to_cpu(hdr->ucode_version);
180 	version_major = (ucode_version >> 20) & 0xfff;
181 	version_minor = (ucode_version >> 8) & 0xfff;
182 	binary_id = ucode_version & 0xff;
183 	DRM_INFO("Found VCE firmware Version: %d.%d Binary ID: %d\n",
184 		version_major, version_minor, binary_id);
185 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
186 				(binary_id << 8));
187 
188 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
189 				    AMDGPU_GEM_DOMAIN_VRAM |
190 				    AMDGPU_GEM_DOMAIN_GTT,
191 				    &adev->vce.vcpu_bo,
192 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
193 	if (r) {
194 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
195 		return r;
196 	}
197 
198 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
199 		atomic_set(&adev->vce.handles[i], 0);
200 		adev->vce.filp[i] = NULL;
201 	}
202 
203 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
204 	mutex_init(&adev->vce.idle_mutex);
205 
206 	return 0;
207 }
208 
209 /**
210  * amdgpu_vce_sw_fini - free memory
211  *
212  * @adev: amdgpu_device pointer
213  *
214  * Last step on VCE teardown, free firmware memory
215  */
216 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
217 {
218 	unsigned i;
219 
220 	if (adev->vce.vcpu_bo == NULL)
221 		return 0;
222 
223 	drm_sched_entity_destroy(&adev->vce.entity);
224 
225 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
226 		(void **)&adev->vce.cpu_addr);
227 
228 	for (i = 0; i < adev->vce.num_rings; i++)
229 		amdgpu_ring_fini(&adev->vce.ring[i]);
230 
231 	release_firmware(adev->vce.fw);
232 	mutex_destroy(&adev->vce.idle_mutex);
233 
234 	return 0;
235 }
236 
237 /**
238  * amdgpu_vce_entity_init - init entity
239  *
240  * @adev: amdgpu_device pointer
241  *
242  */
243 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
244 {
245 	struct amdgpu_ring *ring;
246 	struct drm_gpu_scheduler *sched;
247 	int r;
248 
249 	ring = &adev->vce.ring[0];
250 	sched = &ring->sched;
251 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
252 				  &sched, 1, NULL);
253 	if (r != 0) {
254 		DRM_ERROR("Failed setting up VCE run queue.\n");
255 		return r;
256 	}
257 
258 	return 0;
259 }
260 
261 /**
262  * amdgpu_vce_suspend - unpin VCE fw memory
263  *
264  * @adev: amdgpu_device pointer
265  *
266  */
267 int amdgpu_vce_suspend(struct amdgpu_device *adev)
268 {
269 	int i;
270 
271 	cancel_delayed_work_sync(&adev->vce.idle_work);
272 
273 	if (adev->vce.vcpu_bo == NULL)
274 		return 0;
275 
276 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
277 		if (atomic_read(&adev->vce.handles[i]))
278 			break;
279 
280 	if (i == AMDGPU_MAX_VCE_HANDLES)
281 		return 0;
282 
283 	/* TODO: suspending running encoding sessions isn't supported */
284 	return -EINVAL;
285 }
286 
287 /**
288  * amdgpu_vce_resume - pin VCE fw memory
289  *
290  * @adev: amdgpu_device pointer
291  *
292  */
293 int amdgpu_vce_resume(struct amdgpu_device *adev)
294 {
295 	void *cpu_addr;
296 	const struct common_firmware_header *hdr;
297 	unsigned offset;
298 	int r, idx;
299 
300 	if (adev->vce.vcpu_bo == NULL)
301 		return -EINVAL;
302 
303 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
304 	if (r) {
305 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
306 		return r;
307 	}
308 
309 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
310 	if (r) {
311 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
312 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
313 		return r;
314 	}
315 
316 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
317 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
318 
319 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
320 		memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
321 			    adev->vce.fw->size - offset);
322 		drm_dev_exit(idx);
323 	}
324 
325 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
326 
327 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
328 
329 	return 0;
330 }
331 
332 /**
333  * amdgpu_vce_idle_work_handler - power off VCE
334  *
335  * @work: pointer to work structure
336  *
337  * power of VCE when it's not used any more
338  */
339 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
340 {
341 	struct amdgpu_device *adev =
342 		container_of(work, struct amdgpu_device, vce.idle_work.work);
343 	unsigned i, count = 0;
344 
345 	for (i = 0; i < adev->vce.num_rings; i++)
346 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
347 
348 	if (count == 0) {
349 		if (adev->pm.dpm_enabled) {
350 			amdgpu_dpm_enable_vce(adev, false);
351 		} else {
352 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
353 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
354 							       AMD_PG_STATE_GATE);
355 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
356 							       AMD_CG_STATE_GATE);
357 		}
358 	} else {
359 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
360 	}
361 }
362 
363 /**
364  * amdgpu_vce_ring_begin_use - power up VCE
365  *
366  * @ring: amdgpu ring
367  *
368  * Make sure VCE is powerd up when we want to use it
369  */
370 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
371 {
372 	struct amdgpu_device *adev = ring->adev;
373 	bool set_clocks;
374 
375 	if (amdgpu_sriov_vf(adev))
376 		return;
377 
378 	mutex_lock(&adev->vce.idle_mutex);
379 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
380 	if (set_clocks) {
381 		if (adev->pm.dpm_enabled) {
382 			amdgpu_dpm_enable_vce(adev, true);
383 		} else {
384 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
385 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
386 							       AMD_CG_STATE_UNGATE);
387 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
388 							       AMD_PG_STATE_UNGATE);
389 
390 		}
391 	}
392 	mutex_unlock(&adev->vce.idle_mutex);
393 }
394 
395 /**
396  * amdgpu_vce_ring_end_use - power VCE down
397  *
398  * @ring: amdgpu ring
399  *
400  * Schedule work to power VCE down again
401  */
402 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
403 {
404 	if (!amdgpu_sriov_vf(ring->adev))
405 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
406 }
407 
408 /**
409  * amdgpu_vce_free_handles - free still open VCE handles
410  *
411  * @adev: amdgpu_device pointer
412  * @filp: drm file pointer
413  *
414  * Close all VCE handles still open by this file pointer
415  */
416 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
417 {
418 	struct amdgpu_ring *ring = &adev->vce.ring[0];
419 	int i, r;
420 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
421 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
422 
423 		if (!handle || adev->vce.filp[i] != filp)
424 			continue;
425 
426 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
427 		if (r)
428 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
429 
430 		adev->vce.filp[i] = NULL;
431 		atomic_set(&adev->vce.handles[i], 0);
432 	}
433 }
434 
435 /**
436  * amdgpu_vce_get_create_msg - generate a VCE create msg
437  *
438  * @ring: ring we should submit the msg to
439  * @handle: VCE session handle to use
440  * @fence: optional fence to return
441  *
442  * Open up a stream for HW test
443  */
444 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
445 				     struct dma_fence **fence)
446 {
447 	const unsigned ib_size_dw = 1024;
448 	struct amdgpu_job *job;
449 	struct amdgpu_ib *ib;
450 	struct amdgpu_ib ib_msg;
451 	struct dma_fence *f = NULL;
452 	uint64_t addr;
453 	int i, r;
454 
455 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
456 				     AMDGPU_FENCE_OWNER_UNDEFINED,
457 				     ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT,
458 				     &job);
459 	if (r)
460 		return r;
461 
462 	memset(&ib_msg, 0, sizeof(ib_msg));
463 	/* only one gpu page is needed, alloc +1 page to make addr aligned. */
464 	r = amdgpu_ib_get(ring->adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2,
465 			  AMDGPU_IB_POOL_DIRECT,
466 			  &ib_msg);
467 	if (r)
468 		goto err;
469 
470 	ib = &job->ibs[0];
471 	/* let addr point to page boundary */
472 	addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg.gpu_addr);
473 
474 	/* stitch together an VCE create msg */
475 	ib->length_dw = 0;
476 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
477 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
478 	ib->ptr[ib->length_dw++] = handle;
479 
480 	if ((ring->adev->vce.fw_version >> 24) >= 52)
481 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
482 	else
483 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
484 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
485 	ib->ptr[ib->length_dw++] = 0x00000000;
486 	ib->ptr[ib->length_dw++] = 0x00000042;
487 	ib->ptr[ib->length_dw++] = 0x0000000a;
488 	ib->ptr[ib->length_dw++] = 0x00000001;
489 	ib->ptr[ib->length_dw++] = 0x00000080;
490 	ib->ptr[ib->length_dw++] = 0x00000060;
491 	ib->ptr[ib->length_dw++] = 0x00000100;
492 	ib->ptr[ib->length_dw++] = 0x00000100;
493 	ib->ptr[ib->length_dw++] = 0x0000000c;
494 	ib->ptr[ib->length_dw++] = 0x00000000;
495 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
496 		ib->ptr[ib->length_dw++] = 0x00000000;
497 		ib->ptr[ib->length_dw++] = 0x00000000;
498 		ib->ptr[ib->length_dw++] = 0x00000000;
499 		ib->ptr[ib->length_dw++] = 0x00000000;
500 	}
501 
502 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
503 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
504 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
505 	ib->ptr[ib->length_dw++] = addr;
506 	ib->ptr[ib->length_dw++] = 0x00000001;
507 
508 	for (i = ib->length_dw; i < ib_size_dw; ++i)
509 		ib->ptr[i] = 0x0;
510 
511 	r = amdgpu_job_submit_direct(job, ring, &f);
512 	amdgpu_ib_free(ring->adev, &ib_msg, f);
513 	if (r)
514 		goto err;
515 
516 	if (fence)
517 		*fence = dma_fence_get(f);
518 	dma_fence_put(f);
519 	return 0;
520 
521 err:
522 	amdgpu_job_free(job);
523 	return r;
524 }
525 
526 /**
527  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
528  *
529  * @ring: ring we should submit the msg to
530  * @handle: VCE session handle to use
531  * @direct: direct or delayed pool
532  * @fence: optional fence to return
533  *
534  * Close up a stream for HW test or if userspace failed to do so
535  */
536 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
537 				      bool direct, struct dma_fence **fence)
538 {
539 	const unsigned ib_size_dw = 1024;
540 	struct amdgpu_job *job;
541 	struct amdgpu_ib *ib;
542 	struct dma_fence *f = NULL;
543 	int i, r;
544 
545 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
546 				     AMDGPU_FENCE_OWNER_UNDEFINED,
547 				     ib_size_dw * 4,
548 				     direct ? AMDGPU_IB_POOL_DIRECT :
549 				     AMDGPU_IB_POOL_DELAYED, &job);
550 	if (r)
551 		return r;
552 
553 	ib = &job->ibs[0];
554 
555 	/* stitch together an VCE destroy msg */
556 	ib->length_dw = 0;
557 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
558 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
559 	ib->ptr[ib->length_dw++] = handle;
560 
561 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
562 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
563 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
564 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
565 	ib->ptr[ib->length_dw++] = 0x00000000;
566 	ib->ptr[ib->length_dw++] = 0x00000000;
567 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
568 	ib->ptr[ib->length_dw++] = 0x00000000;
569 
570 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
571 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
572 
573 	for (i = ib->length_dw; i < ib_size_dw; ++i)
574 		ib->ptr[i] = 0x0;
575 
576 	if (direct)
577 		r = amdgpu_job_submit_direct(job, ring, &f);
578 	else
579 		f = amdgpu_job_submit(job);
580 	if (r)
581 		goto err;
582 
583 	if (fence)
584 		*fence = dma_fence_get(f);
585 	dma_fence_put(f);
586 	return 0;
587 
588 err:
589 	amdgpu_job_free(job);
590 	return r;
591 }
592 
593 /**
594  * amdgpu_vce_validate_bo - make sure not to cross 4GB boundary
595  *
596  * @ib: indirect buffer to use
597  * @lo: address of lower dword
598  * @hi: address of higher dword
599  * @size: minimum size
600  * @index: bs/fb index
601  *
602  * Make sure that no BO cross a 4GB boundary.
603  */
604 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p,
605 				  struct amdgpu_ib *ib, int lo, int hi,
606 				  unsigned size, int32_t index)
607 {
608 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
609 	struct ttm_operation_ctx ctx = { false, false };
610 	struct amdgpu_bo_va_mapping *mapping;
611 	unsigned i, fpfn, lpfn;
612 	struct amdgpu_bo *bo;
613 	uint64_t addr;
614 	int r;
615 
616 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
617 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
618 	if (index >= 0) {
619 		addr += offset;
620 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
621 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
622 	} else {
623 		fpfn = 0;
624 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
625 	}
626 
627 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
628 	if (r) {
629 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
630 			  addr, lo, hi, size, index);
631 		return r;
632 	}
633 
634 	for (i = 0; i < bo->placement.num_placement; ++i) {
635 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
636 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
637 			min(bo->placements[i].lpfn, lpfn) : lpfn;
638 	}
639 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
640 }
641 
642 
643 /**
644  * amdgpu_vce_cs_reloc - command submission relocation
645  *
646  * @p: parser context
647  * @ib: indirect buffer to use
648  * @lo: address of lower dword
649  * @hi: address of higher dword
650  * @size: minimum size
651  * @index: bs/fb index
652  *
653  * Patch relocation inside command stream with real buffer address
654  */
655 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, struct amdgpu_ib *ib,
656 			       int lo, int hi, unsigned size, uint32_t index)
657 {
658 	struct amdgpu_bo_va_mapping *mapping;
659 	struct amdgpu_bo *bo;
660 	uint64_t addr;
661 	int r;
662 
663 	if (index == 0xffffffff)
664 		index = 0;
665 
666 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
667 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
668 	addr += ((uint64_t)size) * ((uint64_t)index);
669 
670 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
671 	if (r) {
672 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
673 			  addr, lo, hi, size, index);
674 		return r;
675 	}
676 
677 	if ((addr + (uint64_t)size) >
678 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
679 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
680 			  addr, lo, hi);
681 		return -EINVAL;
682 	}
683 
684 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
685 	addr += amdgpu_bo_gpu_offset(bo);
686 	addr -= ((uint64_t)size) * ((uint64_t)index);
687 
688 	amdgpu_ib_set_value(ib, lo, lower_32_bits(addr));
689 	amdgpu_ib_set_value(ib, hi, upper_32_bits(addr));
690 
691 	return 0;
692 }
693 
694 /**
695  * amdgpu_vce_validate_handle - validate stream handle
696  *
697  * @p: parser context
698  * @handle: handle to validate
699  * @allocated: allocated a new handle?
700  *
701  * Validates the handle and return the found session index or -EINVAL
702  * we we don't have another free session index.
703  */
704 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
705 				      uint32_t handle, uint32_t *allocated)
706 {
707 	unsigned i;
708 
709 	/* validate the handle */
710 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
711 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
712 			if (p->adev->vce.filp[i] != p->filp) {
713 				DRM_ERROR("VCE handle collision detected!\n");
714 				return -EINVAL;
715 			}
716 			return i;
717 		}
718 	}
719 
720 	/* handle not found try to alloc a new one */
721 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
722 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
723 			p->adev->vce.filp[i] = p->filp;
724 			p->adev->vce.img_size[i] = 0;
725 			*allocated |= 1 << i;
726 			return i;
727 		}
728 	}
729 
730 	DRM_ERROR("No more free VCE handles!\n");
731 	return -EINVAL;
732 }
733 
734 /**
735  * amdgpu_vce_ring_parse_cs - parse and validate the command stream
736  *
737  * @p: parser context
738  * @job: the job to parse
739  * @ib: the IB to patch
740  */
741 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p,
742 			     struct amdgpu_job *job,
743 			     struct amdgpu_ib *ib)
744 {
745 	unsigned fb_idx = 0, bs_idx = 0;
746 	int session_idx = -1;
747 	uint32_t destroyed = 0;
748 	uint32_t created = 0;
749 	uint32_t allocated = 0;
750 	uint32_t tmp, handle = 0;
751 	uint32_t *size = &tmp;
752 	unsigned idx;
753 	int i, r = 0;
754 
755 	job->vm = NULL;
756 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
757 
758 	for (idx = 0; idx < ib->length_dw;) {
759 		uint32_t len = amdgpu_ib_get_value(ib, idx);
760 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
761 
762 		if ((len < 8) || (len & 3)) {
763 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
764 			r = -EINVAL;
765 			goto out;
766 		}
767 
768 		switch (cmd) {
769 		case 0x00000002: /* task info */
770 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
771 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
772 			break;
773 
774 		case 0x03000001: /* encode */
775 			r = amdgpu_vce_validate_bo(p, ib, idx + 10, idx + 9,
776 						   0, 0);
777 			if (r)
778 				goto out;
779 
780 			r = amdgpu_vce_validate_bo(p, ib, idx + 12, idx + 11,
781 						   0, 0);
782 			if (r)
783 				goto out;
784 			break;
785 
786 		case 0x05000001: /* context buffer */
787 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
788 						   0, 0);
789 			if (r)
790 				goto out;
791 			break;
792 
793 		case 0x05000004: /* video bitstream buffer */
794 			tmp = amdgpu_ib_get_value(ib, idx + 4);
795 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
796 						   tmp, bs_idx);
797 			if (r)
798 				goto out;
799 			break;
800 
801 		case 0x05000005: /* feedback buffer */
802 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
803 						   4096, fb_idx);
804 			if (r)
805 				goto out;
806 			break;
807 
808 		case 0x0500000d: /* MV buffer */
809 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
810 						   0, 0);
811 			if (r)
812 				goto out;
813 
814 			r = amdgpu_vce_validate_bo(p, ib, idx + 8, idx + 7,
815 						   0, 0);
816 			if (r)
817 				goto out;
818 			break;
819 		}
820 
821 		idx += len / 4;
822 	}
823 
824 	for (idx = 0; idx < ib->length_dw;) {
825 		uint32_t len = amdgpu_ib_get_value(ib, idx);
826 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
827 
828 		switch (cmd) {
829 		case 0x00000001: /* session */
830 			handle = amdgpu_ib_get_value(ib, idx + 2);
831 			session_idx = amdgpu_vce_validate_handle(p, handle,
832 								 &allocated);
833 			if (session_idx < 0) {
834 				r = session_idx;
835 				goto out;
836 			}
837 			size = &p->adev->vce.img_size[session_idx];
838 			break;
839 
840 		case 0x00000002: /* task info */
841 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
842 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
843 			break;
844 
845 		case 0x01000001: /* create */
846 			created |= 1 << session_idx;
847 			if (destroyed & (1 << session_idx)) {
848 				destroyed &= ~(1 << session_idx);
849 				allocated |= 1 << session_idx;
850 
851 			} else if (!(allocated & (1 << session_idx))) {
852 				DRM_ERROR("Handle already in use!\n");
853 				r = -EINVAL;
854 				goto out;
855 			}
856 
857 			*size = amdgpu_ib_get_value(ib, idx + 8) *
858 				amdgpu_ib_get_value(ib, idx + 10) *
859 				8 * 3 / 2;
860 			break;
861 
862 		case 0x04000001: /* config extension */
863 		case 0x04000002: /* pic control */
864 		case 0x04000005: /* rate control */
865 		case 0x04000007: /* motion estimation */
866 		case 0x04000008: /* rdo */
867 		case 0x04000009: /* vui */
868 		case 0x05000002: /* auxiliary buffer */
869 		case 0x05000009: /* clock table */
870 			break;
871 
872 		case 0x0500000c: /* hw config */
873 			switch (p->adev->asic_type) {
874 #ifdef CONFIG_DRM_AMDGPU_CIK
875 			case CHIP_KAVERI:
876 			case CHIP_MULLINS:
877 #endif
878 			case CHIP_CARRIZO:
879 				break;
880 			default:
881 				r = -EINVAL;
882 				goto out;
883 			}
884 			break;
885 
886 		case 0x03000001: /* encode */
887 			r = amdgpu_vce_cs_reloc(p, ib, idx + 10, idx + 9,
888 						*size, 0);
889 			if (r)
890 				goto out;
891 
892 			r = amdgpu_vce_cs_reloc(p, ib, idx + 12, idx + 11,
893 						*size / 3, 0);
894 			if (r)
895 				goto out;
896 			break;
897 
898 		case 0x02000001: /* destroy */
899 			destroyed |= 1 << session_idx;
900 			break;
901 
902 		case 0x05000001: /* context buffer */
903 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
904 						*size * 2, 0);
905 			if (r)
906 				goto out;
907 			break;
908 
909 		case 0x05000004: /* video bitstream buffer */
910 			tmp = amdgpu_ib_get_value(ib, idx + 4);
911 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
912 						tmp, bs_idx);
913 			if (r)
914 				goto out;
915 			break;
916 
917 		case 0x05000005: /* feedback buffer */
918 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
919 						4096, fb_idx);
920 			if (r)
921 				goto out;
922 			break;
923 
924 		case 0x0500000d: /* MV buffer */
925 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3,
926 						idx + 2, *size, 0);
927 			if (r)
928 				goto out;
929 
930 			r = amdgpu_vce_cs_reloc(p, ib, idx + 8,
931 						idx + 7, *size / 12, 0);
932 			if (r)
933 				goto out;
934 			break;
935 
936 		default:
937 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
938 			r = -EINVAL;
939 			goto out;
940 		}
941 
942 		if (session_idx == -1) {
943 			DRM_ERROR("no session command at start of IB\n");
944 			r = -EINVAL;
945 			goto out;
946 		}
947 
948 		idx += len / 4;
949 	}
950 
951 	if (allocated & ~created) {
952 		DRM_ERROR("New session without create command!\n");
953 		r = -ENOENT;
954 	}
955 
956 out:
957 	if (!r) {
958 		/* No error, free all destroyed handle slots */
959 		tmp = destroyed;
960 	} else {
961 		/* Error during parsing, free all allocated handle slots */
962 		tmp = allocated;
963 	}
964 
965 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
966 		if (tmp & (1 << i))
967 			atomic_set(&p->adev->vce.handles[i], 0);
968 
969 	return r;
970 }
971 
972 /**
973  * amdgpu_vce_ring_parse_cs_vm - parse the command stream in VM mode
974  *
975  * @p: parser context
976  * @job: the job to parse
977  * @ib: the IB to patch
978  */
979 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p,
980 				struct amdgpu_job *job,
981 				struct amdgpu_ib *ib)
982 {
983 	int session_idx = -1;
984 	uint32_t destroyed = 0;
985 	uint32_t created = 0;
986 	uint32_t allocated = 0;
987 	uint32_t tmp, handle = 0;
988 	int i, r = 0, idx = 0;
989 
990 	while (idx < ib->length_dw) {
991 		uint32_t len = amdgpu_ib_get_value(ib, idx);
992 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
993 
994 		if ((len < 8) || (len & 3)) {
995 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
996 			r = -EINVAL;
997 			goto out;
998 		}
999 
1000 		switch (cmd) {
1001 		case 0x00000001: /* session */
1002 			handle = amdgpu_ib_get_value(ib, idx + 2);
1003 			session_idx = amdgpu_vce_validate_handle(p, handle,
1004 								 &allocated);
1005 			if (session_idx < 0) {
1006 				r = session_idx;
1007 				goto out;
1008 			}
1009 			break;
1010 
1011 		case 0x01000001: /* create */
1012 			created |= 1 << session_idx;
1013 			if (destroyed & (1 << session_idx)) {
1014 				destroyed &= ~(1 << session_idx);
1015 				allocated |= 1 << session_idx;
1016 
1017 			} else if (!(allocated & (1 << session_idx))) {
1018 				DRM_ERROR("Handle already in use!\n");
1019 				r = -EINVAL;
1020 				goto out;
1021 			}
1022 
1023 			break;
1024 
1025 		case 0x02000001: /* destroy */
1026 			destroyed |= 1 << session_idx;
1027 			break;
1028 
1029 		default:
1030 			break;
1031 		}
1032 
1033 		if (session_idx == -1) {
1034 			DRM_ERROR("no session command at start of IB\n");
1035 			r = -EINVAL;
1036 			goto out;
1037 		}
1038 
1039 		idx += len / 4;
1040 	}
1041 
1042 	if (allocated & ~created) {
1043 		DRM_ERROR("New session without create command!\n");
1044 		r = -ENOENT;
1045 	}
1046 
1047 out:
1048 	if (!r) {
1049 		/* No error, free all destroyed handle slots */
1050 		tmp = destroyed;
1051 		amdgpu_ib_free(p->adev, ib, NULL);
1052 	} else {
1053 		/* Error during parsing, free all allocated handle slots */
1054 		tmp = allocated;
1055 	}
1056 
1057 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1058 		if (tmp & (1 << i))
1059 			atomic_set(&p->adev->vce.handles[i], 0);
1060 
1061 	return r;
1062 }
1063 
1064 /**
1065  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1066  *
1067  * @ring: engine to use
1068  * @job: job to retrieve vmid from
1069  * @ib: the IB to execute
1070  * @flags: unused
1071  *
1072  */
1073 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1074 				struct amdgpu_job *job,
1075 				struct amdgpu_ib *ib,
1076 				uint32_t flags)
1077 {
1078 	amdgpu_ring_write(ring, VCE_CMD_IB);
1079 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1080 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1081 	amdgpu_ring_write(ring, ib->length_dw);
1082 }
1083 
1084 /**
1085  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1086  *
1087  * @ring: engine to use
1088  * @addr: address
1089  * @seq: sequence number
1090  * @flags: fence related flags
1091  *
1092  */
1093 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1094 				unsigned flags)
1095 {
1096 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1097 
1098 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1099 	amdgpu_ring_write(ring, addr);
1100 	amdgpu_ring_write(ring, upper_32_bits(addr));
1101 	amdgpu_ring_write(ring, seq);
1102 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1103 	amdgpu_ring_write(ring, VCE_CMD_END);
1104 }
1105 
1106 /**
1107  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1108  *
1109  * @ring: the engine to test on
1110  *
1111  */
1112 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1113 {
1114 	struct amdgpu_device *adev = ring->adev;
1115 	uint32_t rptr;
1116 	unsigned i;
1117 	int r, timeout = adev->usec_timeout;
1118 
1119 	/* skip ring test for sriov*/
1120 	if (amdgpu_sriov_vf(adev))
1121 		return 0;
1122 
1123 	r = amdgpu_ring_alloc(ring, 16);
1124 	if (r)
1125 		return r;
1126 
1127 	rptr = amdgpu_ring_get_rptr(ring);
1128 
1129 	amdgpu_ring_write(ring, VCE_CMD_END);
1130 	amdgpu_ring_commit(ring);
1131 
1132 	for (i = 0; i < timeout; i++) {
1133 		if (amdgpu_ring_get_rptr(ring) != rptr)
1134 			break;
1135 		udelay(1);
1136 	}
1137 
1138 	if (i >= timeout)
1139 		r = -ETIMEDOUT;
1140 
1141 	return r;
1142 }
1143 
1144 /**
1145  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1146  *
1147  * @ring: the engine to test on
1148  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1149  *
1150  */
1151 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1152 {
1153 	struct dma_fence *fence = NULL;
1154 	long r;
1155 
1156 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1157 	if (ring != &ring->adev->vce.ring[0])
1158 		return 0;
1159 
1160 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
1161 	if (r)
1162 		goto error;
1163 
1164 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1165 	if (r)
1166 		goto error;
1167 
1168 	r = dma_fence_wait_timeout(fence, false, timeout);
1169 	if (r == 0)
1170 		r = -ETIMEDOUT;
1171 	else if (r > 0)
1172 		r = 0;
1173 
1174 error:
1175 	dma_fence_put(fence);
1176 	return r;
1177 }
1178 
1179 enum amdgpu_ring_priority_level amdgpu_vce_get_ring_prio(int ring)
1180 {
1181 	switch(ring) {
1182 	case 0:
1183 		return AMDGPU_RING_PRIO_0;
1184 	case 1:
1185 		return AMDGPU_RING_PRIO_1;
1186 	case 2:
1187 		return AMDGPU_RING_PRIO_2;
1188 	default:
1189 		return AMDGPU_RING_PRIO_0;
1190 	}
1191 }
1192