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