xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c (revision 7a5f1cd22d47f8ca4b760b6334378ae42c1bd24b)
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_SI
45 #define FIRMWARE_VCE_V1_0	"amdgpu/vce_1_0_0.bin"
46 #endif
47 #ifdef CONFIG_DRM_AMDGPU_CIK
48 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_vce.bin"
49 #define FIRMWARE_KABINI	"amdgpu/kabini_vce.bin"
50 #define FIRMWARE_KAVERI	"amdgpu/kaveri_vce.bin"
51 #define FIRMWARE_HAWAII	"amdgpu/hawaii_vce.bin"
52 #define FIRMWARE_MULLINS	"amdgpu/mullins_vce.bin"
53 #endif
54 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
55 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
56 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
57 #define FIRMWARE_STONEY		"amdgpu/stoney_vce.bin"
58 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_vce.bin"
59 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_vce.bin"
60 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_vce.bin"
61 #define FIRMWARE_VEGAM		"amdgpu/vegam_vce.bin"
62 
63 #define FIRMWARE_VEGA10		"amdgpu/vega10_vce.bin"
64 #define FIRMWARE_VEGA12		"amdgpu/vega12_vce.bin"
65 #define FIRMWARE_VEGA20		"amdgpu/vega20_vce.bin"
66 
67 #ifdef CONFIG_DRM_AMDGPU_SI
68 MODULE_FIRMWARE(FIRMWARE_VCE_V1_0);
69 #endif
70 #ifdef CONFIG_DRM_AMDGPU_CIK
71 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
72 MODULE_FIRMWARE(FIRMWARE_KABINI);
73 MODULE_FIRMWARE(FIRMWARE_KAVERI);
74 MODULE_FIRMWARE(FIRMWARE_HAWAII);
75 MODULE_FIRMWARE(FIRMWARE_MULLINS);
76 #endif
77 MODULE_FIRMWARE(FIRMWARE_TONGA);
78 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
79 MODULE_FIRMWARE(FIRMWARE_FIJI);
80 MODULE_FIRMWARE(FIRMWARE_STONEY);
81 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
82 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
83 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
84 MODULE_FIRMWARE(FIRMWARE_VEGAM);
85 
86 MODULE_FIRMWARE(FIRMWARE_VEGA10);
87 MODULE_FIRMWARE(FIRMWARE_VEGA12);
88 MODULE_FIRMWARE(FIRMWARE_VEGA20);
89 
90 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
91 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
92 				     struct dma_fence **fence);
93 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
94 				      bool direct, struct dma_fence **fence);
95 
96 /**
97  * amdgpu_vce_firmware_name() - determine the firmware file name for VCE
98  *
99  * @adev: amdgpu_device pointer
100  *
101  * Each chip that has VCE IP may need a different firmware.
102  * This function returns the name of the VCE firmware file
103  * appropriate for the current chip.
104  */
105 static const char *amdgpu_vce_firmware_name(struct amdgpu_device *adev)
106 {
107 	switch (adev->asic_type) {
108 #ifdef CONFIG_DRM_AMDGPU_SI
109 	case CHIP_PITCAIRN:
110 	case CHIP_TAHITI:
111 	case CHIP_VERDE:
112 		return FIRMWARE_VCE_V1_0;
113 #endif
114 #ifdef CONFIG_DRM_AMDGPU_CIK
115 	case CHIP_BONAIRE:
116 		return FIRMWARE_BONAIRE;
117 	case CHIP_KAVERI:
118 		return FIRMWARE_KAVERI;
119 	case CHIP_KABINI:
120 		return FIRMWARE_KABINI;
121 	case CHIP_HAWAII:
122 		return FIRMWARE_HAWAII;
123 	case CHIP_MULLINS:
124 		return FIRMWARE_MULLINS;
125 #endif
126 	case CHIP_TONGA:
127 		return  FIRMWARE_TONGA;
128 	case CHIP_CARRIZO:
129 		return  FIRMWARE_CARRIZO;
130 	case CHIP_FIJI:
131 		return  FIRMWARE_FIJI;
132 	case CHIP_STONEY:
133 		return  FIRMWARE_STONEY;
134 	case CHIP_POLARIS10:
135 		return  FIRMWARE_POLARIS10;
136 	case CHIP_POLARIS11:
137 		return  FIRMWARE_POLARIS11;
138 	case CHIP_POLARIS12:
139 		return  FIRMWARE_POLARIS12;
140 	case CHIP_VEGAM:
141 		return  FIRMWARE_VEGAM;
142 	case CHIP_VEGA10:
143 		return  FIRMWARE_VEGA10;
144 	case CHIP_VEGA12:
145 		return  FIRMWARE_VEGA12;
146 	case CHIP_VEGA20:
147 		return  FIRMWARE_VEGA20;
148 
149 	default:
150 		return NULL;
151 	}
152 }
153 
154 /**
155  * amdgpu_vce_early_init() - try to load VCE firmware
156  *
157  * @adev: amdgpu_device pointer
158  *
159  * Tries to load the VCE firmware.
160  *
161  * When not found, returns ENOENT so that the driver can
162  * still load and initialize the rest of the IP blocks.
163  * The GPU can function just fine without VCE, they will just
164  * not support video encoding.
165  */
166 int amdgpu_vce_early_init(struct amdgpu_device *adev)
167 {
168 	const char *fw_name = amdgpu_vce_firmware_name(adev);
169 	const struct common_firmware_header *hdr;
170 	unsigned int ucode_version, version_major, version_minor, binary_id;
171 	int r;
172 
173 	if (!fw_name)
174 		return -ENOENT;
175 
176 	r = amdgpu_ucode_request(adev, &adev->vce.fw, AMDGPU_UCODE_REQUIRED, "%s", fw_name);
177 	if (r) {
178 		dev_err(adev->dev,
179 			"amdgpu_vce: Firmware \"%s\" not found or failed to validate (%d)\n",
180 			fw_name, r);
181 
182 		amdgpu_ucode_release(&adev->vce.fw);
183 		return -ENOENT;
184 	}
185 
186 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
187 
188 	ucode_version = le32_to_cpu(hdr->ucode_version);
189 	version_major = (ucode_version >> 20) & 0xfff;
190 	version_minor = (ucode_version >> 8) & 0xfff;
191 	binary_id = ucode_version & 0xff;
192 	dev_info(adev->dev, "Found VCE firmware Version: %d.%d Binary ID: %d\n",
193 		version_major, version_minor, binary_id);
194 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
195 				(binary_id << 8));
196 
197 	return 0;
198 }
199 
200 /**
201  * amdgpu_vce_sw_init() - allocate memory for VCE BO
202  *
203  * @adev: amdgpu_device pointer
204  * @size: size for the new BO
205  *
206  * First step to get VCE online: allocate memory for VCE BO.
207  * The VCE firmware binary is copied into the VCE BO later,
208  * in amdgpu_vce_resume. The VCE executes its code from the
209  * VCE BO and also uses the space in this BO for its stack and data.
210  *
211  * Ideally this BO should be placed in VRAM for optimal performance,
212  * although technically it also runs from system RAM (albeit slowly).
213  */
214 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
215 {
216 	int i, r;
217 
218 	if (!adev->vce.fw)
219 		return -ENOENT;
220 
221 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
222 				    AMDGPU_GEM_DOMAIN_VRAM |
223 				    AMDGPU_GEM_DOMAIN_GTT,
224 				    &adev->vce.vcpu_bo,
225 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
226 	if (r) {
227 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
228 		return r;
229 	}
230 
231 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
232 		atomic_set(&adev->vce.handles[i], 0);
233 		adev->vce.filp[i] = NULL;
234 	}
235 
236 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
237 	mutex_init(&adev->vce.idle_mutex);
238 
239 	return 0;
240 }
241 
242 /**
243  * amdgpu_vce_sw_fini - free memory
244  *
245  * @adev: amdgpu_device pointer
246  *
247  * Last step on VCE teardown, free firmware memory
248  */
249 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
250 {
251 	unsigned int i;
252 
253 	if (adev->vce.vcpu_bo == NULL)
254 		return 0;
255 
256 	drm_sched_entity_destroy(&adev->vce.entity);
257 
258 	for (i = 0; i < adev->vce.num_rings; i++)
259 		amdgpu_ring_fini(&adev->vce.ring[i]);
260 
261 	amdgpu_ucode_release(&adev->vce.fw);
262 	mutex_destroy(&adev->vce.idle_mutex);
263 
264 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
265 		(void **)&adev->vce.cpu_addr);
266 
267 	return 0;
268 }
269 
270 /**
271  * amdgpu_vce_entity_init - init entity
272  *
273  * @adev: amdgpu_device pointer
274  * @ring: amdgpu_ring pointer to check
275  *
276  * Initialize the entity used for handle management in the kernel driver.
277  */
278 int amdgpu_vce_entity_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
279 {
280 	if (ring == &adev->vce.ring[0]) {
281 		struct drm_gpu_scheduler *sched = &ring->sched;
282 		int r;
283 
284 		r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
285 					  &sched, 1, NULL);
286 		if (r != 0) {
287 			DRM_ERROR("Failed setting up VCE run queue.\n");
288 			return r;
289 		}
290 	}
291 
292 	return 0;
293 }
294 
295 /**
296  * amdgpu_vce_suspend - unpin VCE fw memory
297  *
298  * @adev: amdgpu_device pointer
299  *
300  */
301 int amdgpu_vce_suspend(struct amdgpu_device *adev)
302 {
303 	int i;
304 
305 	cancel_delayed_work_sync(&adev->vce.idle_work);
306 
307 	if (adev->vce.vcpu_bo == NULL)
308 		return 0;
309 
310 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
311 		if (atomic_read(&adev->vce.handles[i]))
312 			break;
313 
314 	if (i == AMDGPU_MAX_VCE_HANDLES)
315 		return 0;
316 
317 	/* TODO: suspending running encoding sessions isn't supported */
318 	return -EINVAL;
319 }
320 
321 /**
322  * amdgpu_vce_resume - pin VCE fw memory
323  *
324  * @adev: amdgpu_device pointer
325  *
326  */
327 int amdgpu_vce_resume(struct amdgpu_device *adev)
328 {
329 	const struct common_firmware_header *hdr;
330 	unsigned int offset;
331 	int idx;
332 
333 	if (adev->vce.vcpu_bo == NULL)
334 		return -EINVAL;
335 
336 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
337 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
338 
339 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
340 		memset_io(adev->vce.cpu_addr, 0, amdgpu_bo_size(adev->vce.vcpu_bo));
341 		memcpy_toio(adev->vce.cpu_addr, adev->vce.fw->data + offset,
342 			    adev->vce.fw->size - offset);
343 		drm_dev_exit(idx);
344 	}
345 
346 	return 0;
347 }
348 
349 /**
350  * amdgpu_vce_idle_work_handler - power off VCE
351  *
352  * @work: pointer to work structure
353  *
354  * power of VCE when it's not used any more
355  */
356 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
357 {
358 	struct amdgpu_device *adev =
359 		container_of(work, struct amdgpu_device, vce.idle_work.work);
360 	unsigned int i, count = 0;
361 
362 	for (i = 0; i < adev->vce.num_rings; i++)
363 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
364 
365 	if (count == 0) {
366 		if (adev->pm.dpm_enabled) {
367 			amdgpu_dpm_enable_vce(adev, false);
368 		} else {
369 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
370 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
371 							       AMD_PG_STATE_GATE);
372 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
373 							       AMD_CG_STATE_GATE);
374 		}
375 	} else {
376 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
377 	}
378 }
379 
380 /**
381  * amdgpu_vce_ring_begin_use - power up VCE
382  *
383  * @ring: amdgpu ring
384  *
385  * Make sure VCE is powerd up when we want to use it
386  */
387 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
388 {
389 	struct amdgpu_device *adev = ring->adev;
390 	bool set_clocks;
391 
392 	if (amdgpu_sriov_vf(adev))
393 		return;
394 
395 	mutex_lock(&adev->vce.idle_mutex);
396 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
397 	if (set_clocks) {
398 		if (adev->pm.dpm_enabled) {
399 			amdgpu_dpm_enable_vce(adev, true);
400 		} else {
401 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
402 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
403 							       AMD_CG_STATE_UNGATE);
404 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
405 							       AMD_PG_STATE_UNGATE);
406 
407 		}
408 	}
409 	mutex_unlock(&adev->vce.idle_mutex);
410 }
411 
412 /**
413  * amdgpu_vce_ring_end_use - power VCE down
414  *
415  * @ring: amdgpu ring
416  *
417  * Schedule work to power VCE down again
418  */
419 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
420 {
421 	if (!amdgpu_sriov_vf(ring->adev))
422 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
423 }
424 
425 /**
426  * amdgpu_vce_free_handles - free still open VCE handles
427  *
428  * @adev: amdgpu_device pointer
429  * @filp: drm file pointer
430  *
431  * Close all VCE handles still open by this file pointer
432  */
433 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
434 {
435 	struct amdgpu_ring *ring = &adev->vce.ring[0];
436 	int i, r;
437 
438 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
439 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
440 
441 		if (!handle || adev->vce.filp[i] != filp)
442 			continue;
443 
444 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
445 		if (r)
446 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
447 
448 		adev->vce.filp[i] = NULL;
449 		atomic_set(&adev->vce.handles[i], 0);
450 	}
451 }
452 
453 /**
454  * amdgpu_vce_get_create_msg - generate a VCE create msg
455  *
456  * @ring: ring we should submit the msg to
457  * @handle: VCE session handle to use
458  * @fence: optional fence to return
459  *
460  * Open up a stream for HW test
461  */
462 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
463 				     struct dma_fence **fence)
464 {
465 	const unsigned int ib_size_dw = 1024;
466 	struct amdgpu_job *job;
467 	struct amdgpu_ib *ib;
468 	struct amdgpu_ib ib_msg;
469 	struct dma_fence *f = NULL;
470 	uint64_t addr;
471 	int i, r;
472 
473 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
474 				     AMDGPU_FENCE_OWNER_UNDEFINED,
475 				     ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT,
476 				     &job, AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST);
477 	if (r)
478 		return r;
479 
480 	memset(&ib_msg, 0, sizeof(ib_msg));
481 	/* only one gpu page is needed, alloc +1 page to make addr aligned. */
482 	r = amdgpu_ib_get(ring->adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2,
483 			  AMDGPU_IB_POOL_DIRECT,
484 			  &ib_msg);
485 	if (r)
486 		goto err;
487 
488 	ib = &job->ibs[0];
489 	/* let addr point to page boundary */
490 	addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg.gpu_addr);
491 
492 	/* stitch together an VCE create msg */
493 	ib->length_dw = 0;
494 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
495 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
496 	ib->ptr[ib->length_dw++] = handle;
497 
498 	if ((ring->adev->vce.fw_version >> 24) >= 52)
499 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
500 	else
501 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
502 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
503 	ib->ptr[ib->length_dw++] = 0x00000000;
504 	ib->ptr[ib->length_dw++] = 0x00000042;
505 	ib->ptr[ib->length_dw++] = 0x0000000a;
506 	ib->ptr[ib->length_dw++] = 0x00000001;
507 	ib->ptr[ib->length_dw++] = 0x00000080;
508 	ib->ptr[ib->length_dw++] = 0x00000060;
509 	ib->ptr[ib->length_dw++] = 0x00000100;
510 	ib->ptr[ib->length_dw++] = 0x00000100;
511 	ib->ptr[ib->length_dw++] = 0x0000000c;
512 	ib->ptr[ib->length_dw++] = 0x00000000;
513 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
514 		ib->ptr[ib->length_dw++] = 0x00000000;
515 		ib->ptr[ib->length_dw++] = 0x00000000;
516 		ib->ptr[ib->length_dw++] = 0x00000000;
517 		ib->ptr[ib->length_dw++] = 0x00000000;
518 	}
519 
520 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
521 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
522 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
523 	ib->ptr[ib->length_dw++] = addr;
524 	ib->ptr[ib->length_dw++] = 0x00000001;
525 
526 	for (i = ib->length_dw; i < ib_size_dw; ++i)
527 		ib->ptr[i] = 0x0;
528 
529 	r = amdgpu_job_submit_direct(job, ring, &f);
530 	amdgpu_ib_free(&ib_msg, f);
531 	if (r)
532 		goto err;
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_get_destroy_msg - generate a VCE destroy msg
546  *
547  * @ring: ring we should submit the msg to
548  * @handle: VCE session handle to use
549  * @direct: direct or delayed pool
550  * @fence: optional fence to return
551  *
552  * Close up a stream for HW test or if userspace failed to do so
553  */
554 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
555 				      bool direct, struct dma_fence **fence)
556 {
557 	const unsigned int ib_size_dw = 1024;
558 	struct amdgpu_job *job;
559 	struct amdgpu_ib *ib;
560 	struct dma_fence *f = NULL;
561 	int i, r;
562 
563 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
564 				     AMDGPU_FENCE_OWNER_UNDEFINED,
565 				     ib_size_dw * 4,
566 				     direct ? AMDGPU_IB_POOL_DIRECT :
567 				     AMDGPU_IB_POOL_DELAYED, &job,
568 				     AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST);
569 	if (r)
570 		return r;
571 
572 	ib = &job->ibs[0];
573 
574 	/* stitch together an VCE destroy msg */
575 	ib->length_dw = 0;
576 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
577 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
578 	ib->ptr[ib->length_dw++] = handle;
579 
580 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
581 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
582 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
583 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
584 	ib->ptr[ib->length_dw++] = 0x00000000;
585 	ib->ptr[ib->length_dw++] = 0x00000000;
586 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
587 	ib->ptr[ib->length_dw++] = 0x00000000;
588 
589 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
590 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
591 
592 	for (i = ib->length_dw; i < ib_size_dw; ++i)
593 		ib->ptr[i] = 0x0;
594 
595 	if (direct)
596 		r = amdgpu_job_submit_direct(job, ring, &f);
597 	else
598 		f = amdgpu_job_submit(job);
599 	if (r)
600 		goto err;
601 
602 	if (fence)
603 		*fence = dma_fence_get(f);
604 	dma_fence_put(f);
605 	return 0;
606 
607 err:
608 	amdgpu_job_free(job);
609 	return r;
610 }
611 
612 /**
613  * amdgpu_vce_validate_bo - make sure not to cross 4GB boundary
614  *
615  * @p: cs parser
616  * @ib: indirect buffer to use
617  * @lo: address of lower dword
618  * @hi: address of higher dword
619  * @size: minimum size
620  * @index: bs/fb index
621  *
622  * Make sure that no BO cross a 4GB boundary.
623  */
624 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p,
625 				  struct amdgpu_ib *ib, int lo, int hi,
626 				  unsigned int size, int32_t index)
627 {
628 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
629 	struct ttm_operation_ctx ctx = { false, false };
630 	struct amdgpu_bo_va_mapping *mapping;
631 	unsigned int i, fpfn, lpfn;
632 	struct amdgpu_bo *bo;
633 	uint64_t addr;
634 	int r;
635 
636 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
637 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
638 	if (index >= 0) {
639 		addr += offset;
640 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
641 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
642 	} else {
643 		fpfn = 0;
644 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
645 	}
646 
647 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
648 	if (r) {
649 		DRM_ERROR("Can't find BO for addr 0x%010llx %d %d %d %d\n",
650 			  addr, lo, hi, size, index);
651 		return r;
652 	}
653 
654 	for (i = 0; i < bo->placement.num_placement; ++i) {
655 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
656 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
657 			min(bo->placements[i].lpfn, lpfn) : lpfn;
658 	}
659 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
660 }
661 
662 
663 /**
664  * amdgpu_vce_cs_reloc - command submission relocation
665  *
666  * @p: parser context
667  * @ib: indirect buffer to use
668  * @lo: address of lower dword
669  * @hi: address of higher dword
670  * @size: minimum size
671  * @index: bs/fb index
672  *
673  * Patch relocation inside command stream with real buffer address
674  */
675 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, struct amdgpu_ib *ib,
676 			       int lo, int hi, unsigned int size, uint32_t index)
677 {
678 	struct amdgpu_bo_va_mapping *mapping;
679 	struct amdgpu_bo *bo;
680 	uint64_t addr;
681 	int r;
682 
683 	if (lo >= ib->length_dw || hi >= ib->length_dw)
684 		return -EINVAL;
685 
686 	if (index == 0xffffffff)
687 		index = 0;
688 
689 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
690 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
691 	addr += ((uint64_t)size) * ((uint64_t)index);
692 
693 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
694 	if (r) {
695 		DRM_ERROR("Can't find BO for addr 0x%010llx %d %d %d %d\n",
696 			  addr, lo, hi, size, index);
697 		return r;
698 	}
699 
700 	if ((addr + (uint64_t)size) >
701 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
702 		DRM_ERROR("BO too small for addr 0x%010llx %d %d\n",
703 			  addr, lo, hi);
704 		return -EINVAL;
705 	}
706 
707 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
708 	addr += amdgpu_bo_gpu_offset(bo);
709 	addr -= ((uint64_t)size) * ((uint64_t)index);
710 
711 	amdgpu_ib_set_value(ib, lo, lower_32_bits(addr));
712 	amdgpu_ib_set_value(ib, hi, upper_32_bits(addr));
713 
714 	return 0;
715 }
716 
717 /**
718  * amdgpu_vce_validate_handle - validate stream handle
719  *
720  * @p: parser context
721  * @handle: handle to validate
722  * @allocated: allocated a new handle?
723  *
724  * Validates the handle and return the found session index or -EINVAL
725  * we don't have another free session index.
726  */
727 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
728 				      uint32_t handle, uint32_t *allocated)
729 {
730 	unsigned int i;
731 
732 	/* validate the handle */
733 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
734 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
735 			if (p->adev->vce.filp[i] != p->filp) {
736 				DRM_ERROR("VCE handle collision detected!\n");
737 				return -EINVAL;
738 			}
739 			return i;
740 		}
741 	}
742 
743 	/* handle not found try to alloc a new one */
744 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
745 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
746 			p->adev->vce.filp[i] = p->filp;
747 			p->adev->vce.img_size[i] = 0;
748 			*allocated |= 1 << i;
749 			return i;
750 		}
751 	}
752 
753 	DRM_ERROR("No more free VCE handles!\n");
754 	return -EINVAL;
755 }
756 
757 /**
758  * amdgpu_vce_ring_parse_cs - parse and validate the command stream
759  *
760  * @p: parser context
761  * @job: the job to parse
762  * @ib: the IB to patch
763  */
764 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p,
765 			     struct amdgpu_job *job,
766 			     struct amdgpu_ib *ib)
767 {
768 	unsigned int fb_idx = 0, bs_idx = 0;
769 	int session_idx = -1;
770 	uint32_t destroyed = 0;
771 	uint32_t created = 0;
772 	uint32_t allocated = 0;
773 	uint32_t tmp, handle = 0;
774 	uint32_t dummy = 0xffffffff;
775 	uint32_t *size = &dummy;
776 	unsigned int idx;
777 	int i, r = 0;
778 
779 	job->vm = NULL;
780 
781 	for (idx = 0; idx < ib->length_dw;) {
782 		uint32_t len = amdgpu_ib_get_value(ib, idx);
783 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
784 
785 		if ((len < 8) || (len & 3)) {
786 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
787 			r = -EINVAL;
788 			goto out;
789 		}
790 
791 		switch (cmd) {
792 		case 0x00000002: /* task info */
793 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
794 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
795 			break;
796 
797 		case 0x03000001: /* encode */
798 			r = amdgpu_vce_validate_bo(p, ib, idx + 10, idx + 9,
799 						   0, 0);
800 			if (r)
801 				goto out;
802 
803 			r = amdgpu_vce_validate_bo(p, ib, idx + 12, idx + 11,
804 						   0, 0);
805 			if (r)
806 				goto out;
807 			break;
808 
809 		case 0x05000001: /* context buffer */
810 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
811 						   0, 0);
812 			if (r)
813 				goto out;
814 			break;
815 
816 		case 0x05000004: /* video bitstream buffer */
817 			tmp = amdgpu_ib_get_value(ib, idx + 4);
818 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
819 						   tmp, bs_idx);
820 			if (r)
821 				goto out;
822 			break;
823 
824 		case 0x05000005: /* feedback buffer */
825 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
826 						   4096, fb_idx);
827 			if (r)
828 				goto out;
829 			break;
830 
831 		case 0x0500000d: /* MV buffer */
832 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
833 						   0, 0);
834 			if (r)
835 				goto out;
836 
837 			r = amdgpu_vce_validate_bo(p, ib, idx + 8, idx + 7,
838 						   0, 0);
839 			if (r)
840 				goto out;
841 			break;
842 		}
843 
844 		idx += len / 4;
845 	}
846 
847 	for (idx = 0; idx < ib->length_dw;) {
848 		uint32_t len = amdgpu_ib_get_value(ib, idx);
849 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
850 
851 		switch (cmd) {
852 		case 0x00000001: /* session */
853 			handle = amdgpu_ib_get_value(ib, idx + 2);
854 			session_idx = amdgpu_vce_validate_handle(p, handle,
855 								 &allocated);
856 			if (session_idx < 0) {
857 				r = session_idx;
858 				goto out;
859 			}
860 			size = &p->adev->vce.img_size[session_idx];
861 			break;
862 
863 		case 0x00000002: /* task info */
864 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
865 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
866 			break;
867 
868 		case 0x01000001: /* create */
869 			created |= 1 << session_idx;
870 			if (destroyed & (1 << session_idx)) {
871 				destroyed &= ~(1 << session_idx);
872 				allocated |= 1 << session_idx;
873 
874 			} else if (!(allocated & (1 << session_idx))) {
875 				DRM_ERROR("Handle already in use!\n");
876 				r = -EINVAL;
877 				goto out;
878 			}
879 
880 			*size = amdgpu_ib_get_value(ib, idx + 8) *
881 				amdgpu_ib_get_value(ib, idx + 10) *
882 				8 * 3 / 2;
883 			break;
884 
885 		case 0x04000001: /* config extension */
886 		case 0x04000002: /* pic control */
887 		case 0x04000005: /* rate control */
888 		case 0x04000007: /* motion estimation */
889 		case 0x04000008: /* rdo */
890 		case 0x04000009: /* vui */
891 		case 0x05000002: /* auxiliary buffer */
892 		case 0x05000009: /* clock table */
893 			break;
894 
895 		case 0x0500000c: /* hw config */
896 			switch (p->adev->asic_type) {
897 #ifdef CONFIG_DRM_AMDGPU_CIK
898 			case CHIP_KAVERI:
899 			case CHIP_MULLINS:
900 #endif
901 			case CHIP_CARRIZO:
902 				break;
903 			default:
904 				r = -EINVAL;
905 				goto out;
906 			}
907 			break;
908 
909 		case 0x03000001: /* encode */
910 			r = amdgpu_vce_cs_reloc(p, ib, idx + 10, idx + 9,
911 						*size, 0);
912 			if (r)
913 				goto out;
914 
915 			r = amdgpu_vce_cs_reloc(p, ib, idx + 12, idx + 11,
916 						*size / 3, 0);
917 			if (r)
918 				goto out;
919 			break;
920 
921 		case 0x02000001: /* destroy */
922 			destroyed |= 1 << session_idx;
923 			break;
924 
925 		case 0x05000001: /* context buffer */
926 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
927 						*size * 2, 0);
928 			if (r)
929 				goto out;
930 			break;
931 
932 		case 0x05000004: /* video bitstream buffer */
933 			tmp = amdgpu_ib_get_value(ib, idx + 4);
934 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
935 						tmp, bs_idx);
936 			if (r)
937 				goto out;
938 			break;
939 
940 		case 0x05000005: /* feedback buffer */
941 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
942 						4096, fb_idx);
943 			if (r)
944 				goto out;
945 			break;
946 
947 		case 0x0500000d: /* MV buffer */
948 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3,
949 						idx + 2, *size, 0);
950 			if (r)
951 				goto out;
952 
953 			r = amdgpu_vce_cs_reloc(p, ib, idx + 8,
954 						idx + 7, *size / 12, 0);
955 			if (r)
956 				goto out;
957 			break;
958 
959 		default:
960 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
961 			r = -EINVAL;
962 			goto out;
963 		}
964 
965 		if (session_idx == -1) {
966 			DRM_ERROR("no session command at start of IB\n");
967 			r = -EINVAL;
968 			goto out;
969 		}
970 
971 		idx += len / 4;
972 	}
973 
974 	if (allocated & ~created) {
975 		DRM_ERROR("New session without create command!\n");
976 		r = -ENOENT;
977 	}
978 
979 out:
980 	if (!r) {
981 		/* No error, free all destroyed handle slots */
982 		tmp = destroyed;
983 	} else {
984 		/* Error during parsing, free all allocated handle slots */
985 		tmp = allocated;
986 	}
987 
988 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
989 		if (tmp & (1 << i))
990 			atomic_set(&p->adev->vce.handles[i], 0);
991 
992 	return r;
993 }
994 
995 /**
996  * amdgpu_vce_ring_parse_cs_vm - parse the command stream in VM mode
997  *
998  * @p: parser context
999  * @job: the job to parse
1000  * @ib: the IB to patch
1001  */
1002 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p,
1003 				struct amdgpu_job *job,
1004 				struct amdgpu_ib *ib)
1005 {
1006 	int session_idx = -1;
1007 	uint32_t destroyed = 0;
1008 	uint32_t created = 0;
1009 	uint32_t allocated = 0;
1010 	uint32_t tmp, handle = 0;
1011 	int i, r = 0, idx = 0;
1012 
1013 	while (idx < ib->length_dw) {
1014 		uint32_t len = amdgpu_ib_get_value(ib, idx);
1015 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
1016 
1017 		if ((len < 8) || (len & 3)) {
1018 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
1019 			r = -EINVAL;
1020 			goto out;
1021 		}
1022 
1023 		switch (cmd) {
1024 		case 0x00000001: /* session */
1025 			handle = amdgpu_ib_get_value(ib, idx + 2);
1026 			session_idx = amdgpu_vce_validate_handle(p, handle,
1027 								 &allocated);
1028 			if (session_idx < 0) {
1029 				r = session_idx;
1030 				goto out;
1031 			}
1032 			break;
1033 
1034 		case 0x01000001: /* create */
1035 			created |= 1 << session_idx;
1036 			if (destroyed & (1 << session_idx)) {
1037 				destroyed &= ~(1 << session_idx);
1038 				allocated |= 1 << session_idx;
1039 
1040 			} else if (!(allocated & (1 << session_idx))) {
1041 				DRM_ERROR("Handle already in use!\n");
1042 				r = -EINVAL;
1043 				goto out;
1044 			}
1045 
1046 			break;
1047 
1048 		case 0x02000001: /* destroy */
1049 			destroyed |= 1 << session_idx;
1050 			break;
1051 
1052 		default:
1053 			break;
1054 		}
1055 
1056 		if (session_idx == -1) {
1057 			DRM_ERROR("no session command at start of IB\n");
1058 			r = -EINVAL;
1059 			goto out;
1060 		}
1061 
1062 		idx += len / 4;
1063 	}
1064 
1065 	if (allocated & ~created) {
1066 		DRM_ERROR("New session without create command!\n");
1067 		r = -ENOENT;
1068 	}
1069 
1070 out:
1071 	if (!r) {
1072 		/* No error, free all destroyed handle slots */
1073 		tmp = destroyed;
1074 	} else {
1075 		/* Error during parsing, free all allocated handle slots */
1076 		tmp = allocated;
1077 	}
1078 
1079 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1080 		if (tmp & (1 << i))
1081 			atomic_set(&p->adev->vce.handles[i], 0);
1082 
1083 	return r;
1084 }
1085 
1086 /**
1087  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1088  *
1089  * @ring: engine to use
1090  * @job: job to retrieve vmid from
1091  * @ib: the IB to execute
1092  * @flags: unused
1093  *
1094  */
1095 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1096 				struct amdgpu_job *job,
1097 				struct amdgpu_ib *ib,
1098 				uint32_t flags)
1099 {
1100 	amdgpu_ring_write(ring, VCE_CMD_IB);
1101 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1102 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1103 	amdgpu_ring_write(ring, ib->length_dw);
1104 }
1105 
1106 /**
1107  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1108  *
1109  * @ring: engine to use
1110  * @addr: address
1111  * @seq: sequence number
1112  * @flags: fence related flags
1113  *
1114  */
1115 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1116 				unsigned int flags)
1117 {
1118 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1119 
1120 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1121 	amdgpu_ring_write(ring, addr);
1122 	amdgpu_ring_write(ring, upper_32_bits(addr));
1123 	amdgpu_ring_write(ring, seq);
1124 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1125 	amdgpu_ring_write(ring, VCE_CMD_END);
1126 }
1127 
1128 /**
1129  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1130  *
1131  * @ring: the engine to test on
1132  *
1133  */
1134 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1135 {
1136 	struct amdgpu_device *adev = ring->adev;
1137 	uint32_t rptr;
1138 	unsigned int i;
1139 	int r, timeout = adev->usec_timeout;
1140 
1141 	/* skip ring test for sriov*/
1142 	if (amdgpu_sriov_vf(adev))
1143 		return 0;
1144 
1145 	r = amdgpu_ring_alloc(ring, 16);
1146 	if (r)
1147 		return r;
1148 
1149 	rptr = amdgpu_ring_get_rptr(ring);
1150 
1151 	amdgpu_ring_write(ring, VCE_CMD_END);
1152 	amdgpu_ring_commit(ring);
1153 
1154 	for (i = 0; i < timeout; i++) {
1155 		if (amdgpu_ring_get_rptr(ring) != rptr)
1156 			break;
1157 		udelay(1);
1158 	}
1159 
1160 	if (i >= timeout)
1161 		r = -ETIMEDOUT;
1162 
1163 	return r;
1164 }
1165 
1166 /**
1167  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1168  *
1169  * @ring: the engine to test on
1170  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1171  *
1172  */
1173 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1174 {
1175 	struct dma_fence *fence = NULL;
1176 	long r;
1177 
1178 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1179 	if (ring != &ring->adev->vce.ring[0])
1180 		return 0;
1181 
1182 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
1183 	if (r)
1184 		goto error;
1185 
1186 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1187 	if (r)
1188 		goto error;
1189 
1190 	r = dma_fence_wait_timeout(fence, false, timeout);
1191 	if (r == 0)
1192 		r = -ETIMEDOUT;
1193 	else if (r > 0)
1194 		r = 0;
1195 
1196 error:
1197 	dma_fence_put(fence);
1198 	return r;
1199 }
1200 
1201 enum amdgpu_ring_priority_level amdgpu_vce_get_ring_prio(int ring)
1202 {
1203 	switch (ring) {
1204 	case 0:
1205 		return AMDGPU_RING_PRIO_0;
1206 	case 1:
1207 		return AMDGPU_RING_PRIO_1;
1208 	case 2:
1209 		return AMDGPU_RING_PRIO_2;
1210 	default:
1211 		return AMDGPU_RING_PRIO_0;
1212 	}
1213 }
1214